{"schema_version":"1.0","gateway":{"key":"higress","display":"Higress","lang":"Other","cls":"API gateway","repo":"https://github.com/alibaba/higress","version":"2.2.5","stars":9764,"stars_as_of":"2026-11-14","first_commit":"2022-10-27","matrix":{"gateway":"higress","build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","matrix_version":null,"served":true,"serve_error":"","upstream_shape":"","upstream_note":"","egress_configured":"","probe_first":false,"capability_note":"","cell_perf_sweep":true,"sweep_rung_selection":null,"sweep_ttft_ms":null,"p99_ceiling_ms":null,"sweep_dur":null,"cell_stream":true,"cell_memory":true,"memory":null,"cells":{},"upstreams":{"cohere":{"configurable":true,"served":true,"egress_config":null,"serve_error":"","cells":{"cohere":{"served":true,"reason":null,"status":"","path":"/v2/chat","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":615,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":747,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":774,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":27,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":613.2194813409235,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1630.737493553378,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.8053044884180178,"certified":true,"suppressed":false},"cost_window_ok":{"value":31620,"certified":true,"suppressed":false},"cost_window_rps":{"value":5269,"certified":true,"suppressed":false},"cost_threads":{"value":143,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.0076217583807716635,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":3092,"certified":true,"suppressed":false},"concurrency":2,"p99_us":790,"first_disqualified_conc":4,"lower_bound":false,"samples":18555},{"bound_ms":5,"rps":{"value":5921,"certified":true,"suppressed":false},"concurrency":8,"p99_us":2939,"first_disqualified_conc":16,"lower_bound":false,"samples":35536},{"bound_ms":10,"rps":{"value":6529,"certified":true,"suppressed":false},"concurrency":32,"p99_us":9093,"first_disqualified_conc":64,"lower_bound":false,"samples":39189},{"bound_ms":50,"rps":{"value":6633,"certified":true,"suppressed":false},"concurrency":64,"p99_us":18236,"first_disqualified_conc":256,"lower_bound":false,"samples":39805},{"bound_ms":100,"rps":{"value":6633,"certified":true,"suppressed":false},"concurrency":64,"p99_us":18236,"first_disqualified_conc":512,"lower_bound":false,"samples":39805},{"bound_ms":null,"rps":{"value":6633,"certified":true,"suppressed":false},"concurrency":64,"p99_us":18236,"first_disqualified_conc":null,"lower_bound":false,"samples":39805}],"sweep":[{"conc":1,"ok":9281,"rps":1546,"p99_us":819,"fail":0},{"conc":1,"ok":9245,"rps":1540,"p99_us":836,"fail":0},{"conc":1,"ok":9269,"rps":1544,"p99_us":806,"fail":0},{"conc":2,"ok":18390,"rps":3064,"p99_us":793,"fail":0},{"conc":2,"ok":10460,"rps":1742,"p99_us":1435,"fail":0},{"conc":2,"ok":18555,"rps":3092,"p99_us":790,"fail":0},{"conc":4,"ok":20413,"rps":3401,"p99_us":1475,"fail":0},{"conc":4,"ok":28729,"rps":4787,"p99_us":1381,"fail":0},{"conc":4,"ok":32981,"rps":5495,"p99_us":1263,"fail":0},{"conc":8,"ok":35230,"rps":5870,"p99_us":2496,"fail":0},{"conc":8,"ok":35536,"rps":5921,"p99_us":2939,"fail":0},{"conc":8,"ok":29911,"rps":4984,"p99_us":2636,"fail":0},{"conc":16,"ok":37803,"rps":6299,"p99_us":6458,"fail":0},{"conc":16,"ok":38085,"rps":6346,"p99_us":5752,"fail":0},{"conc":16,"ok":37018,"rps":6168,"p99_us":5602,"fail":0},{"conc":32,"ok":39094,"rps":6514,"p99_us":10998,"fail":0},{"conc":32,"ok":39171,"rps":6527,"p99_us":10605,"fail":0},{"conc":32,"ok":39189,"rps":6529,"p99_us":9093,"fail":0},{"conc":64,"ok":39793,"rps":6631,"p99_us":17089,"fail":0},{"conc":64,"ok":39806,"rps":6633,"p99_us":17057,"fail":0},{"conc":64,"ok":39805,"rps":6633,"p99_us":18236,"fail":0},{"conc":128,"ok":39670,"rps":6610,"p99_us":32755,"fail":0},{"conc":128,"ok":39750,"rps":6623,"p99_us":28708,"fail":0},{"conc":128,"ok":39727,"rps":6620,"p99_us":31453,"fail":0},{"conc":256,"ok":37991,"rps":6331,"p99_us":58947,"fail":0},{"conc":256,"ok":38052,"rps":6340,"p99_us":59566,"fail":0},{"conc":256,"ok":37451,"rps":6240,"p99_us":58802,"fail":0},{"conc":512,"ok":34958,"rps":5824,"p99_us":115106,"fail":0},{"conc":512,"ok":34963,"rps":5825,"p99_us":115301,"fail":0},{"conc":512,"ok":34828,"rps":5802,"p99_us":111254,"fail":0},{"conc":1024,"ok":34718,"rps":5785,"p99_us":838880,"fail":0},{"conc":1024,"ok":35746,"rps":5956,"p99_us":228885,"fail":0},{"conc":1024,"ok":36117,"rps":6018,"p99_us":227755,"fail":0},{"conc":2048,"ok":36159,"rps":6025,"p99_us":750428,"fail":0},{"conc":2048,"ok":35996,"rps":5998,"p99_us":634125,"fail":0},{"conc":2048,"ok":36132,"rps":6021,"p99_us":787241,"fail":0},{"conc":4096,"ok":36355,"rps":6057,"p99_us":855678,"fail":0},{"conc":4096,"ok":36396,"rps":6065,"p99_us":757334,"fail":0},{"conc":4096,"ok":36514,"rps":6085,"p99_us":777264,"fail":0},{"conc":8192,"ok":36115,"rps":6018,"p99_us":612409,"fail":0},{"conc":8192,"ok":36072,"rps":6011,"p99_us":642450,"fail":0},{"conc":8192,"ok":36139,"rps":6022,"p99_us":722649,"fail":0},{"conc":16384,"ok":36390,"rps":6064,"p99_us":547156,"fail":0},{"conc":16384,"ok":36334,"rps":6054,"p99_us":635217,"fail":0},{"conc":16384,"ok":36404,"rps":6066,"p99_us":636580,"fail":0},{"conc":32768,"ok":27417,"rps":4568,"p99_us":2985638,"fail":0},{"conc":32768,"ok":32081,"rps":5345,"p99_us":2616501,"fail":0},{"conc":32768,"ok":32461,"rps":5408,"p99_us":2629504,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1544,"certified":true,"suppressed":false},"p99_us":819,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":3064,"certified":true,"suppressed":false},"p99_us":793,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":4787,"certified":true,"suppressed":false},"p99_us":1381,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":5870,"certified":true,"suppressed":false},"p99_us":2636,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":6299,"certified":true,"suppressed":false},"p99_us":5752,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":6527,"certified":true,"suppressed":false},"p99_us":10605,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":6633,"certified":true,"suppressed":false},"p99_us":17089,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":6620,"certified":true,"suppressed":false},"p99_us":31453,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":6331,"certified":true,"suppressed":false},"p99_us":58947,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":5824,"certified":true,"suppressed":false},"p99_us":115106,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":5956,"certified":true,"suppressed":false},"p99_us":228885,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":6021,"certified":true,"suppressed":false},"p99_us":750428,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":6065,"certified":true,"suppressed":false},"p99_us":777264,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":6018,"certified":true,"suppressed":false},"p99_us":642450,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":6064,"certified":true,"suppressed":false},"p99_us":635217,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":5345,"certified":true,"suppressed":false},"p99_us":2629504,"windows":3,"clean_windows":3}],"reverify_note":"cohere>cohere is a same-dialect cell: there is no translation to prove, so a request-shape check upstream cannot distinguish a translating gateway from a forwarding one","c1_note":"the c=1 percentiles are taken over 9341 successful gateway round trip(s) and 267984 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":9341,"direct_c1_samples":267984},"stream":{"stream_served":"untestable","stream_error":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here","reason":"untestable","stream_c1_note":null,"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"}},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1207.94921875,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1217.2890625,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1218.421875,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1226.8046875,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1324.41015625,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.7482121208635604,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":1.0318152244070045,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1208.60546875},{"t_s":2,"rss_mib":1225.44921875},{"t_s":3,"rss_mib":1225.46484375},{"t_s":5,"rss_mib":1225.48828125},{"t_s":6,"rss_mib":1225.4921875},{"t_s":8,"rss_mib":1225.69921875},{"t_s":9,"rss_mib":1225.69921875},{"t_s":11,"rss_mib":1225.75},{"t_s":12,"rss_mib":1225.75},{"t_s":14,"rss_mib":1225.85546875},{"t_s":15,"rss_mib":1225.85546875},{"t_s":17,"rss_mib":1225.89453125},{"t_s":18,"rss_mib":1225.89453125},{"t_s":19,"rss_mib":1225.9453125},{"t_s":21,"rss_mib":1225.9453125},{"t_s":23,"rss_mib":1226.0078125},{"t_s":24,"rss_mib":1226.01953125},{"t_s":25,"rss_mib":1226.08984375},{"t_s":27,"rss_mib":1226.09765625},{"t_s":29,"rss_mib":1226.1328125},{"t_s":30,"rss_mib":1226.1328125},{"t_s":32,"rss_mib":1226.19921875},{"t_s":33,"rss_mib":1226.19921875},{"t_s":35,"rss_mib":1226.23046875},{"t_s":36,"rss_mib":1226.23046875},{"t_s":38,"rss_mib":1226.30078125},{"t_s":39,"rss_mib":1226.30078125},{"t_s":41,"rss_mib":1226.3515625},{"t_s":42,"rss_mib":1226.3515625},{"t_s":44,"rss_mib":1226.38671875},{"t_s":45,"rss_mib":1226.38671875},{"t_s":47,"rss_mib":1226.453125},{"t_s":48,"rss_mib":1226.60546875},{"t_s":49,"rss_mib":1225.421875},{"t_s":51,"rss_mib":1225.65625},{"t_s":53,"rss_mib":1225.7421875},{"t_s":54,"rss_mib":1225.7421875},{"t_s":55,"rss_mib":1225.8046875},{"t_s":57,"rss_mib":1225.8046875},{"t_s":59,"rss_mib":1225.89453125},{"t_s":60,"rss_mib":1225.89453125},{"t_s":62,"rss_mib":1225.96484375},{"t_s":63,"rss_mib":1225.96484375},{"t_s":64,"rss_mib":1225.9765625},{"t_s":66,"rss_mib":1225.9765625},{"t_s":67,"rss_mib":1226.0390625},{"t_s":69,"rss_mib":1226.0390625},{"t_s":71,"rss_mib":1226.046875},{"t_s":72,"rss_mib":1226.046875},{"t_s":75,"rss_mib":1226.046875},{"t_s":78,"rss_mib":1226.046875},{"t_s":80,"rss_mib":1226.0546875},{"t_s":81,"rss_mib":1226.0546875},{"t_s":82,"rss_mib":1226.0625},{"t_s":84,"rss_mib":1226.0625},{"t_s":85,"rss_mib":1226.0703125},{"t_s":87,"rss_mib":1226.0703125},{"t_s":89,"rss_mib":1226.08984375},{"t_s":90,"rss_mib":1226.08984375},{"t_s":92,"rss_mib":1226.10546875},{"t_s":93,"rss_mib":1226.10546875},{"t_s":95,"rss_mib":1226.12109375},{"t_s":96,"rss_mib":1226.12109375},{"t_s":99,"rss_mib":1226.12109375},{"t_s":101,"rss_mib":1226.14453125},{"t_s":102,"rss_mib":1226.14453125},{"t_s":104,"rss_mib":1226.16796875},{"t_s":105,"rss_mib":1226.16796875},{"t_s":108,"rss_mib":1226.18359375},{"t_s":108,"rss_mib":1226.19921875},{"t_s":110,"rss_mib":1226.296875},{"t_s":111,"rss_mib":1226.296875},{"t_s":113,"rss_mib":1226.41796875},{"t_s":114,"rss_mib":1226.41796875},{"t_s":116,"rss_mib":1226.44921875},{"t_s":117,"rss_mib":1226.44921875},{"t_s":119,"rss_mib":1226.46875},{"t_s":120,"rss_mib":1226.47265625},{"t_s":121,"rss_mib":1226.4921875},{"t_s":123,"rss_mib":1226.49609375},{"t_s":125,"rss_mib":1226.51171875},{"t_s":126,"rss_mib":1226.51171875},{"t_s":128,"rss_mib":1226.53125},{"t_s":129,"rss_mib":1226.53125},{"t_s":130,"rss_mib":1226.5546875},{"t_s":132,"rss_mib":1226.55859375},{"t_s":135,"rss_mib":1226.60546875},{"t_s":135,"rss_mib":1226.61328125},{"t_s":137,"rss_mib":1226.62890625},{"t_s":138,"rss_mib":1226.62890625},{"t_s":140,"rss_mib":1226.64453125},{"t_s":141,"rss_mib":1226.64453125},{"t_s":143,"rss_mib":1226.6796875},{"t_s":144,"rss_mib":1226.6953125},{"t_s":145,"rss_mib":1226.70703125},{"t_s":147,"rss_mib":1226.70703125},{"t_s":148,"rss_mib":1226.71875},{"t_s":150,"rss_mib":1226.71875},{"t_s":150,"rss_mib":1226.72265625},{"t_s":153,"rss_mib":1226.72265625},{"t_s":156,"rss_mib":1226.72265625},{"t_s":159,"rss_mib":1226.72265625},{"t_s":161,"rss_mib":1226.7578125},{"t_s":162,"rss_mib":1226.7578125},{"t_s":163,"rss_mib":1226.77734375},{"t_s":165,"rss_mib":1226.78515625},{"t_s":167,"rss_mib":1226.8046875},{"t_s":168,"rss_mib":1222.20703125},{"t_s":169,"rss_mib":1222.27734375},{"t_s":173,"rss_mib":1222.3359375},{"t_s":174,"rss_mib":1219.55078125},{"t_s":174,"rss_mib":1219.5546875},{"t_s":179,"rss_mib":1219.55859375},{"t_s":179,"rss_mib":1218.18359375},{"t_s":180,"rss_mib":1218.20703125},{"t_s":181,"rss_mib":1218.21484375},{"t_s":183,"rss_mib":1218.21484375},{"t_s":186,"rss_mib":1218.21484375},{"t_s":188,"rss_mib":1218.2578125},{"t_s":189,"rss_mib":1218.2578125},{"t_s":189,"rss_mib":1216.5546875},{"t_s":192,"rss_mib":1216.58203125},{"t_s":195,"rss_mib":1216.5859375},{"t_s":195,"rss_mib":1216.5859375},{"t_s":198,"rss_mib":1216.5859375},{"t_s":201,"rss_mib":1216.5859375},{"t_s":203,"rss_mib":1216.59765625},{"t_s":204,"rss_mib":1216.59765625},{"t_s":205,"rss_mib":1216.60546875},{"t_s":207,"rss_mib":1216.60546875},{"t_s":209,"rss_mib":1216.625},{"t_s":210,"rss_mib":1216.625},{"t_s":212,"rss_mib":1216.640625},{"t_s":213,"rss_mib":1216.640625},{"t_s":214,"rss_mib":1216.6484375},{"t_s":216,"rss_mib":1216.6484375},{"t_s":218,"rss_mib":1216.66796875},{"t_s":219,"rss_mib":1216.66796875},{"t_s":221,"rss_mib":1216.6953125},{"t_s":222,"rss_mib":1216.6953125},{"t_s":225,"rss_mib":1216.7109375},{"t_s":225,"rss_mib":1216.7109375},{"t_s":227,"rss_mib":1216.7265625},{"t_s":228,"rss_mib":1216.7265625},{"t_s":231,"rss_mib":1216.90625},{"t_s":231,"rss_mib":1216.90625},{"t_s":233,"rss_mib":1217.01171875},{"t_s":234,"rss_mib":1217.01953125},{"t_s":237,"rss_mib":1217.0703125},{"t_s":237,"rss_mib":1217.0703125},{"t_s":238,"rss_mib":1217.078125},{"t_s":240,"rss_mib":1217.09375},{"t_s":243,"rss_mib":1217.1640625},{"t_s":243,"rss_mib":1217.1640625},{"t_s":246,"rss_mib":1217.1953125},{"t_s":246,"rss_mib":1217.19921875},{"t_s":247,"rss_mib":1217.20703125},{"t_s":249,"rss_mib":1217.20703125},{"t_s":252,"rss_mib":1217.22265625},{"t_s":252,"rss_mib":1217.22265625},{"t_s":253,"rss_mib":1217.23046875},{"t_s":255,"rss_mib":1217.23046875},{"t_s":257,"rss_mib":1217.24609375},{"t_s":258,"rss_mib":1217.26171875},{"t_s":260,"rss_mib":1217.26953125},{"t_s":261,"rss_mib":1217.26953125},{"t_s":264,"rss_mib":1217.2890625},{"t_s":264,"rss_mib":1217.2890625},{"t_s":267,"rss_mib":1217.30859375},{"t_s":267,"rss_mib":1217.30859375},{"t_s":269,"rss_mib":1217.3203125},{"t_s":270,"rss_mib":1217.328125},{"t_s":273,"rss_mib":1217.3515625},{"t_s":273,"rss_mib":1217.359375},{"t_s":275,"rss_mib":1217.3671875},{"t_s":276,"rss_mib":1217.3671875},{"t_s":279,"rss_mib":1217.3828125},{"t_s":279,"rss_mib":1217.3828125},{"t_s":281,"rss_mib":1217.40625},{"t_s":282,"rss_mib":1217.40625},{"t_s":284,"rss_mib":1217.4296875},{"t_s":285,"rss_mib":1217.4296875},{"t_s":288,"rss_mib":1217.51171875},{"t_s":288,"rss_mib":1217.51171875},{"t_s":289,"rss_mib":1218.95703125},{"t_s":291,"rss_mib":1217.58984375},{"t_s":293,"rss_mib":1217.7109375},{"t_s":294,"rss_mib":1217.7109375},{"t_s":296,"rss_mib":1218.10546875},{"t_s":297,"rss_mib":1218.10546875},{"t_s":299,"rss_mib":1218.11328125},{"t_s":300,"rss_mib":1218.11328125},{"t_s":301,"rss_mib":1218.1171875},{"t_s":303,"rss_mib":1218.1171875},{"t_s":306,"rss_mib":1218.1171875},{"t_s":307,"rss_mib":1218.125},{"t_s":309,"rss_mib":1218.125},{"t_s":311,"rss_mib":1218.13671875},{"t_s":312,"rss_mib":1218.13671875},{"t_s":315,"rss_mib":1218.13671875},{"t_s":317,"rss_mib":1218.1640625},{"t_s":318,"rss_mib":1218.1640625},{"t_s":319,"rss_mib":1218.19921875},{"t_s":321,"rss_mib":1218.19921875},{"t_s":324,"rss_mib":1218.19921875},{"t_s":324,"rss_mib":1218.20703125},{"t_s":327,"rss_mib":1218.20703125},{"t_s":328,"rss_mib":1218.2109375},{"t_s":330,"rss_mib":1218.421875},{"t_s":333,"rss_mib":1218.421875},{"t_s":336,"rss_mib":1218.421875},{"t_s":339,"rss_mib":1218.421875},{"t_s":342,"rss_mib":1218.421875},{"t_s":345,"rss_mib":1218.421875},{"t_s":348,"rss_mib":1218.421875},{"t_s":349,"rss_mib":1219.23828125},{"t_s":351,"rss_mib":1219.2421875},{"t_s":353,"rss_mib":1219.32421875},{"t_s":354,"rss_mib":1219.32421875},{"t_s":357,"rss_mib":1219.32421875},{"t_s":360,"rss_mib":1219.33984375}],"idle_rss_series":[{"t_s":0,"rss_mib":1207.2109375},{"t_s":0,"rss_mib":1207.2734375},{"t_s":0,"rss_mib":1207.2734375},{"t_s":0,"rss_mib":1207.2734375},{"t_s":1,"rss_mib":1207.2890625},{"t_s":1,"rss_mib":1207.2890625},{"t_s":1,"rss_mib":1207.40625},{"t_s":2,"rss_mib":1207.4140625},{"t_s":2,"rss_mib":1207.421875},{"t_s":2,"rss_mib":1207.421875},{"t_s":3,"rss_mib":1207.43359375},{"t_s":3,"rss_mib":1207.43359375},{"t_s":3,"rss_mib":1207.43359375},{"t_s":4,"rss_mib":1207.62890625},{"t_s":4,"rss_mib":1207.62890625},{"t_s":4,"rss_mib":1207.63671875},{"t_s":4,"rss_mib":1207.64453125},{"t_s":5,"rss_mib":1207.6484375},{"t_s":5,"rss_mib":1207.6484375},{"t_s":5,"rss_mib":1207.6484375},{"t_s":6,"rss_mib":1207.65625},{"t_s":6,"rss_mib":1207.65625},{"t_s":6,"rss_mib":1207.6640625},{"t_s":6,"rss_mib":1207.6640625},{"t_s":7,"rss_mib":1207.671875},{"t_s":7,"rss_mib":1207.671875},{"t_s":8,"rss_mib":1207.671875},{"t_s":8,"rss_mib":1207.671875},{"t_s":9,"rss_mib":1207.6875},{"t_s":9,"rss_mib":1207.6875},{"t_s":10,"rss_mib":1207.6875},{"t_s":10,"rss_mib":1207.6875},{"t_s":10,"rss_mib":1207.6953125},{"t_s":10,"rss_mib":1207.6953125},{"t_s":11,"rss_mib":1207.703125},{"t_s":11,"rss_mib":1207.703125},{"t_s":11,"rss_mib":1207.70703125},{"t_s":12,"rss_mib":1207.70703125},{"t_s":12,"rss_mib":1207.7265625},{"t_s":13,"rss_mib":1207.7265625},{"t_s":13,"rss_mib":1207.734375},{"t_s":13,"rss_mib":1207.734375},{"t_s":13,"rss_mib":1207.734375},{"t_s":14,"rss_mib":1207.734375},{"t_s":14,"rss_mib":1207.7421875},{"t_s":14,"rss_mib":1207.7421875},{"t_s":15,"rss_mib":1207.74609375},{"t_s":15,"rss_mib":1207.74609375},{"t_s":15,"rss_mib":1207.74609375},{"t_s":16,"rss_mib":1207.74609375},{"t_s":16,"rss_mib":1207.74609375},{"t_s":17,"rss_mib":1207.75390625},{"t_s":17,"rss_mib":1207.76171875},{"t_s":17,"rss_mib":1207.76171875},{"t_s":18,"rss_mib":1207.76171875},{"t_s":18,"rss_mib":1207.76171875},{"t_s":19,"rss_mib":1207.77734375},{"t_s":19,"rss_mib":1207.77734375},{"t_s":20,"rss_mib":1207.77734375},{"t_s":20,"rss_mib":1207.77734375},{"t_s":20,"rss_mib":1207.78515625},{"t_s":20,"rss_mib":1207.79296875},{"t_s":21,"rss_mib":1207.796875},{"t_s":21,"rss_mib":1207.796875},{"t_s":21,"rss_mib":1207.796875},{"t_s":22,"rss_mib":1207.796875},{"t_s":22,"rss_mib":1207.8046875},{"t_s":23,"rss_mib":1207.8125},{"t_s":23,"rss_mib":1207.8125},{"t_s":23,"rss_mib":1207.8125},{"t_s":24,"rss_mib":1207.859375},{"t_s":24,"rss_mib":1207.859375},{"t_s":24,"rss_mib":1207.8671875},{"t_s":24,"rss_mib":1207.8671875},{"t_s":25,"rss_mib":1207.8828125},{"t_s":25,"rss_mib":1207.8828125},{"t_s":25,"rss_mib":1207.890625},{"t_s":26,"rss_mib":1207.890625},{"t_s":26,"rss_mib":1207.890625},{"t_s":27,"rss_mib":1207.8984375},{"t_s":27,"rss_mib":1207.90625},{"t_s":27,"rss_mib":1207.9140625},{"t_s":27,"rss_mib":1207.9140625},{"t_s":28,"rss_mib":1207.9140625},{"t_s":28,"rss_mib":1207.9140625},{"t_s":29,"rss_mib":1207.94140625},{"t_s":29,"rss_mib":1207.94140625},{"t_s":29,"rss_mib":1207.94921875},{"t_s":30,"rss_mib":1207.94921875},{"t_s":30,"rss_mib":1207.95703125},{"t_s":30,"rss_mib":1207.95703125},{"t_s":30,"rss_mib":1207.95703125},{"t_s":31,"rss_mib":1207.96484375},{"t_s":31,"rss_mib":1207.96484375},{"t_s":31,"rss_mib":1207.96484375},{"t_s":32,"rss_mib":1207.97265625},{"t_s":32,"rss_mib":1207.98046875},{"t_s":33,"rss_mib":1207.98046875},{"t_s":33,"rss_mib":1207.984375},{"t_s":33,"rss_mib":1207.984375},{"t_s":33,"rss_mib":1207.984375},{"t_s":34,"rss_mib":1207.9921875},{"t_s":34,"rss_mib":1208},{"t_s":34,"rss_mib":1208.0078125},{"t_s":35,"rss_mib":1208.01171875},{"t_s":35,"rss_mib":1208.01171875},{"t_s":35,"rss_mib":1208.01953125},{"t_s":36,"rss_mib":1208.01953125},{"t_s":36,"rss_mib":1208.02734375},{"t_s":36,"rss_mib":1208.02734375},{"t_s":36,"rss_mib":1208.03515625},{"t_s":36,"rss_mib":1208.03515625},{"t_s":37,"rss_mib":1208.03515625},{"t_s":37,"rss_mib":1208.04296875},{"t_s":37,"rss_mib":1208.04296875},{"t_s":38,"rss_mib":1208.05078125},{"t_s":38,"rss_mib":1208.05078125},{"t_s":38,"rss_mib":1208.05078125},{"t_s":39,"rss_mib":1208.0546875},{"t_s":39,"rss_mib":1208.0546875},{"t_s":39,"rss_mib":1208.0625},{"t_s":40,"rss_mib":1208.0625},{"t_s":40,"rss_mib":1208.0625},{"t_s":40,"rss_mib":1208.0703125},{"t_s":40,"rss_mib":1208.0703125},{"t_s":41,"rss_mib":1208.0703125},{"t_s":41,"rss_mib":1208.0859375},{"t_s":41,"rss_mib":1208.0859375},{"t_s":42,"rss_mib":1208.0859375},{"t_s":43,"rss_mib":1208.0859375},{"t_s":43,"rss_mib":1208.09375},{"t_s":43,"rss_mib":1208.09375},{"t_s":43,"rss_mib":1208.1015625},{"t_s":43,"rss_mib":1208.1015625},{"t_s":44,"rss_mib":1208.109375},{"t_s":44,"rss_mib":1208.1171875},{"t_s":45,"rss_mib":1208.125},{"t_s":45,"rss_mib":1208.125},{"t_s":45,"rss_mib":1208.1328125},{"t_s":46,"rss_mib":1208.1328125},{"t_s":46,"rss_mib":1208.1328125},{"t_s":46,"rss_mib":1208.1328125},{"t_s":47,"rss_mib":1208.14453125},{"t_s":47,"rss_mib":1208.14453125},{"t_s":47,"rss_mib":1208.15234375},{"t_s":47,"rss_mib":1208.15234375},{"t_s":48,"rss_mib":1208.15234375},{"t_s":48,"rss_mib":1208.16015625},{"t_s":49,"rss_mib":1208.16015625},{"t_s":49,"rss_mib":1208.29296875},{"t_s":49,"rss_mib":1208.29296875},{"t_s":49,"rss_mib":1208.30078125},{"t_s":49,"rss_mib":1208.30078125},{"t_s":50,"rss_mib":1208.3046875},{"t_s":50,"rss_mib":1208.3046875},{"t_s":50,"rss_mib":1208.3125},{"t_s":50,"rss_mib":1208.3125},{"t_s":51,"rss_mib":1208.3125},{"t_s":51,"rss_mib":1208.33203125},{"t_s":51,"rss_mib":1208.33203125},{"t_s":52,"rss_mib":1208.33984375},{"t_s":52,"rss_mib":1208.34375},{"t_s":53,"rss_mib":1208.34375},{"t_s":53,"rss_mib":1208.3515625},{"t_s":53,"rss_mib":1208.484375},{"t_s":53,"rss_mib":1208.4921875},{"t_s":53,"rss_mib":1208.5},{"t_s":54,"rss_mib":1208.50390625},{"t_s":54,"rss_mib":1208.50390625},{"t_s":54,"rss_mib":1208.50390625},{"t_s":55,"rss_mib":1208.51953125},{"t_s":55,"rss_mib":1208.51953125},{"t_s":55,"rss_mib":1208.52734375},{"t_s":56,"rss_mib":1208.52734375},{"t_s":56,"rss_mib":1208.5390625},{"t_s":56,"rss_mib":1208.5390625},{"t_s":56,"rss_mib":1208.5390625},{"t_s":57,"rss_mib":1208.5390625},{"t_s":57,"rss_mib":1208.546875},{"t_s":58,"rss_mib":1208.55859375},{"t_s":58,"rss_mib":1208.55859375},{"t_s":59,"rss_mib":1208.56640625},{"t_s":59,"rss_mib":1208.56640625},{"t_s":59,"rss_mib":1208.56640625}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.009432675,"cost":6.034215506,"memory":434.940222237,"streaming":0.000001879,"streams_sustained":0.000002016,"throughput":321.088546636},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"stream.added_gap_p50_us":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_gap_p99_us":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p50_us":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p99_us":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_direct_samples":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_gw_samples":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_fps":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_headroom":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_mock_ceiling":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_direct_samples":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_gw_samples":{"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here"}}},"bedrock":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/model/command-r-plus/converse","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"gemini":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1beta/models/command-r-plus:generateContent","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"openai":{"served":"failed","reason":"failed","status":"200","path":"/v1/chat/completions","verdict_note":"probed and answered HTTP 200, but the body is not a usable openai response (see body_snippet)","body_snippet":"the 2xx body is not a response of the openai dialect with assistant text at choices[].message.content: {\"response_id\":\"x\",\"generation_id\":\"x\",\"text\":\"ok\",\"finish_reason\":\"COMPLETE\",\"meta\":{\"api_version\":{\"version\":\"1\"},\"tokens\":{\"input_tokens\":10,\"output_tokens\":2},\"billed_units\":{\"input_tokens\":10,\"output_tokens\":2}}}","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"anthropic":{"served":"failed","reason":"failed","status":"200","path":"/v1/messages","verdict_note":"probed and answered HTTP 200, but the body is not a usable anthropic response (see body_snippet)","body_snippet":"the 2xx body is not a response of the anthropic dialect with assistant text at content[type=text].text: {\"id\":\"\",\"type\":\"message\",\"role\":\"assistant\",\"content\":null,\"model\":\"\",\"stop_reason\":null,\"stop_sequence\":null,\"usage\":{},\"error\":null}","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"openai-responses":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1/responses","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}}}},"anthropic":{"configurable":true,"served":true,"egress_config":null,"serve_error":"","cells":{"cohere":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v2/chat","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"gemini":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1beta/models/claude-3-5-sonnet-20241022:generateContent","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"openai":{"served":true,"reason":null,"status":"","path":"/v1/chat/completions","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":859,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":1112,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":1140,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":882.4409213281484,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1133.2203389830509,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.980004881341195,"certified":true,"suppressed":false},"cost_window_ok":{"value":26744,"certified":true,"suppressed":false},"cost_window_rps":{"value":4456,"certified":true,"suppressed":false},"cost_threads":{"value":145,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.001121746933891714,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":4788,"certified":true,"suppressed":false},"concurrency":16,"p99_us":4880,"first_disqualified_conc":32,"lower_bound":false,"samples":28736},{"bound_ms":10,"rps":{"value":4909,"certified":true,"suppressed":false},"concurrency":32,"p99_us":9079,"first_disqualified_conc":64,"lower_bound":false,"samples":29464},{"bound_ms":50,"rps":{"value":4944,"certified":true,"suppressed":false},"concurrency":64,"p99_us":19849,"first_disqualified_conc":256,"lower_bound":false,"samples":29673},{"bound_ms":100,"rps":{"value":4944,"certified":true,"suppressed":false},"concurrency":64,"p99_us":19849,"first_disqualified_conc":512,"lower_bound":false,"samples":29673},{"bound_ms":null,"rps":{"value":4944,"certified":true,"suppressed":false},"concurrency":64,"p99_us":19849,"first_disqualified_conc":null,"lower_bound":false,"samples":29673}],"sweep":[{"conc":1,"ok":6662,"rps":1110,"p99_us":1165,"fail":0},{"conc":1,"ok":6671,"rps":1111,"p99_us":1166,"fail":0},{"conc":1,"ok":6785,"rps":1130,"p99_us":1133,"fail":0},{"conc":2,"ok":12938,"rps":2155,"p99_us":1171,"fail":0},{"conc":2,"ok":12705,"rps":2117,"p99_us":1176,"fail":0},{"conc":2,"ok":12841,"rps":2139,"p99_us":1176,"fail":0},{"conc":4,"ok":13747,"rps":2290,"p99_us":3082,"fail":0},{"conc":4,"ok":20477,"rps":3412,"p99_us":1981,"fail":0},{"conc":4,"ok":26332,"rps":4387,"p99_us":1225,"fail":0},{"conc":8,"ok":27475,"rps":4578,"p99_us":4112,"fail":0},{"conc":8,"ok":27950,"rps":4657,"p99_us":2973,"fail":0},{"conc":8,"ok":27927,"rps":4653,"p99_us":3024,"fail":0},{"conc":16,"ok":28736,"rps":4788,"p99_us":4880,"fail":0},{"conc":16,"ok":28806,"rps":4800,"p99_us":5046,"fail":0},{"conc":16,"ok":28816,"rps":4802,"p99_us":5836,"fail":0},{"conc":32,"ok":29464,"rps":4909,"p99_us":9079,"fail":0},{"conc":32,"ok":29366,"rps":4894,"p99_us":11175,"fail":0},{"conc":32,"ok":29443,"rps":4906,"p99_us":10943,"fail":0},{"conc":64,"ok":29673,"rps":4944,"p99_us":19849,"fail":0},{"conc":64,"ok":29649,"rps":4939,"p99_us":18297,"fail":0},{"conc":64,"ok":29487,"rps":4914,"p99_us":18823,"fail":0},{"conc":128,"ok":29380,"rps":4895,"p99_us":37219,"fail":0},{"conc":128,"ok":29457,"rps":4908,"p99_us":35728,"fail":0},{"conc":128,"ok":29448,"rps":4907,"p99_us":34166,"fail":0},{"conc":256,"ok":27758,"rps":4625,"p99_us":73977,"fail":0},{"conc":256,"ok":27774,"rps":4627,"p99_us":71715,"fail":0},{"conc":256,"ok":27781,"rps":4629,"p99_us":75422,"fail":0},{"conc":512,"ok":27235,"rps":4538,"p99_us":147749,"fail":0},{"conc":512,"ok":27237,"rps":4539,"p99_us":151100,"fail":0},{"conc":512,"ok":27085,"rps":4514,"p99_us":145713,"fail":0},{"conc":1024,"ok":26932,"rps":4487,"p99_us":1877868,"fail":0},{"conc":1024,"ok":27228,"rps":4537,"p99_us":497956,"fail":0},{"conc":1024,"ok":27403,"rps":4566,"p99_us":1259616,"fail":0},{"conc":2048,"ok":26743,"rps":4456,"p99_us":1080307,"fail":0},{"conc":2048,"ok":26981,"rps":4496,"p99_us":1118256,"fail":0},{"conc":2048,"ok":26969,"rps":4494,"p99_us":902839,"fail":0},{"conc":4096,"ok":26965,"rps":4493,"p99_us":1029012,"fail":0},{"conc":4096,"ok":27017,"rps":4501,"p99_us":1007242,"fail":0},{"conc":4096,"ok":26705,"rps":4449,"p99_us":816316,"fail":0},{"conc":8192,"ok":26756,"rps":4459,"p99_us":890982,"fail":0},{"conc":8192,"ok":26783,"rps":4463,"p99_us":1048045,"fail":0},{"conc":8192,"ok":27035,"rps":4504,"p99_us":1013440,"fail":0},{"conc":16384,"ok":27078,"rps":4512,"p99_us":834839,"fail":0},{"conc":16384,"ok":27256,"rps":4542,"p99_us":863669,"fail":0},{"conc":16384,"ok":26890,"rps":4480,"p99_us":758820,"fail":0},{"conc":32768,"ok":21394,"rps":3565,"p99_us":2922297,"fail":0},{"conc":32768,"ok":23790,"rps":3964,"p99_us":2570015,"fail":0},{"conc":32768,"ok":19677,"rps":3279,"p99_us":2759758,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1111,"certified":true,"suppressed":false},"p99_us":1165,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":2139,"certified":true,"suppressed":false},"p99_us":1176,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":3412,"certified":true,"suppressed":false},"p99_us":1981,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":4653,"certified":true,"suppressed":false},"p99_us":3024,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":4800,"certified":true,"suppressed":false},"p99_us":5046,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":4906,"certified":true,"suppressed":false},"p99_us":10943,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":4939,"certified":true,"suppressed":false},"p99_us":18823,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":4907,"certified":true,"suppressed":false},"p99_us":35728,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":4627,"certified":true,"suppressed":false},"p99_us":73977,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":4538,"certified":true,"suppressed":false},"p99_us":147749,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":4537,"certified":true,"suppressed":false},"p99_us":1259616,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":4494,"certified":true,"suppressed":false},"p99_us":1080307,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":4493,"certified":true,"suppressed":false},"p99_us":1007242,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":4463,"certified":true,"suppressed":false},"p99_us":1013440,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":4512,"certified":true,"suppressed":false},"p99_us":834839,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":3565,"certified":true,"suppressed":false},"p99_us":2759758,"windows":3,"clean_windows":3}],"egress_reverified":true,"c1_note":"the c=1 percentiles are taken over 6739 successful gateway round trip(s) and 262868 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":6739,"direct_c1_samples":262868},"stream":{"stream_served":true,"stream_error":null,"reason":null,"stream_c1_note":"the c=1 streaming legs read 64 frame(s) through the gateway and 64 direct to the mock, out of a 64 frame budget. The added-GAP figures are percentiles over the intervals inside those two streams. The added-TTFT figures are percentiles over separate probes - 100 through the gateway and 100 direct, out of 100 attempted per leg - so they carry that weight and no more","added_ttft_p50_us":{"value":1095,"certified":true,"suppressed":false},"added_ttft_p99_us":{"value":1556,"certified":true,"suppressed":false},"added_gap_p50_us":{"value":8,"certified":true,"suppressed":false},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"below_resolution","detail":"the gateway's own inter-frame gap at this percentile (21173us) came in under the mock's (21174us), which a proxy cannot really do - the added gap here is below what this rig can resolve against the mock's 20ms pacing"},"streams_sustained_fps":{"value":23656.559777267295,"certified":true,"suppressed":false,"headroom":0.08075923714495402,"rig_ceiling":292926.98412698414},"streams_sustained":{"value":5767,"certified":true,"suppressed":false,"headroom":0.08075923714495402,"rig_ceiling":292926.98412698414},"ttft_gw_samples":100,"ttft_direct_samples":100,"gap_gw_samples":63,"gap_direct_samples":63},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1148.421875,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1158.875,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1159.7734375,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1167.30859375,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1264.890625,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.7021128067897227,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":0.8890053062912313,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1148.94921875},{"t_s":2,"rss_mib":1165.94140625},{"t_s":3,"rss_mib":1165.94140625},{"t_s":5,"rss_mib":1166.03125},{"t_s":6,"rss_mib":1166.046875},{"t_s":8,"rss_mib":1166.3671875},{"t_s":9,"rss_mib":1166.3671875},{"t_s":11,"rss_mib":1166.42578125},{"t_s":12,"rss_mib":1166.42578125},{"t_s":13,"rss_mib":1166.4609375},{"t_s":15,"rss_mib":1166.4609375},{"t_s":17,"rss_mib":1166.5},{"t_s":18,"rss_mib":1166.5},{"t_s":20,"rss_mib":1166.546875},{"t_s":21,"rss_mib":1166.546875},{"t_s":23,"rss_mib":1166.625},{"t_s":24,"rss_mib":1166.625},{"t_s":26,"rss_mib":1166.71484375},{"t_s":27,"rss_mib":1166.71484375},{"t_s":29,"rss_mib":1166.75},{"t_s":30,"rss_mib":1166.75},{"t_s":32,"rss_mib":1166.8203125},{"t_s":33,"rss_mib":1166.8203125},{"t_s":35,"rss_mib":1166.84765625},{"t_s":36,"rss_mib":1166.84765625},{"t_s":38,"rss_mib":1167.00390625},{"t_s":39,"rss_mib":1167.00390625},{"t_s":41,"rss_mib":1167.0390625},{"t_s":42,"rss_mib":1167.0390625},{"t_s":44,"rss_mib":1167.0703125},{"t_s":45,"rss_mib":1167.0703125},{"t_s":47,"rss_mib":1167.1015625},{"t_s":48,"rss_mib":1167.1015625},{"t_s":49,"rss_mib":1167.23046875},{"t_s":51,"rss_mib":1167.23046875},{"t_s":53,"rss_mib":1167.28125},{"t_s":54,"rss_mib":1167.2890625},{"t_s":54,"rss_mib":1165.83203125},{"t_s":57,"rss_mib":1165.96484375},{"t_s":59,"rss_mib":1166.07421875},{"t_s":60,"rss_mib":1166.07421875},{"t_s":62,"rss_mib":1166.13671875},{"t_s":63,"rss_mib":1166.13671875},{"t_s":65,"rss_mib":1166.140625},{"t_s":66,"rss_mib":1166.140625},{"t_s":66,"rss_mib":1166.15234375},{"t_s":69,"rss_mib":1166.15234375},{"t_s":71,"rss_mib":1166.15625},{"t_s":72,"rss_mib":1166.1796875},{"t_s":72,"rss_mib":1166.1875},{"t_s":75,"rss_mib":1166.1875},{"t_s":78,"rss_mib":1166.1953125},{"t_s":78,"rss_mib":1166.296875},{"t_s":79,"rss_mib":1166.328125},{"t_s":81,"rss_mib":1166.328125},{"t_s":83,"rss_mib":1166.41796875},{"t_s":84,"rss_mib":1166.41796875},{"t_s":85,"rss_mib":1166.4609375},{"t_s":87,"rss_mib":1166.4609375},{"t_s":87,"rss_mib":1166.4765625},{"t_s":90,"rss_mib":1166.4765625},{"t_s":92,"rss_mib":1166.4921875},{"t_s":93,"rss_mib":1166.4921875},{"t_s":94,"rss_mib":1166.5},{"t_s":96,"rss_mib":1166.5},{"t_s":98,"rss_mib":1166.515625},{"t_s":99,"rss_mib":1166.515625},{"t_s":102,"rss_mib":1166.52734375},{"t_s":102,"rss_mib":1166.52734375},{"t_s":104,"rss_mib":1166.5390625},{"t_s":105,"rss_mib":1166.5390625},{"t_s":106,"rss_mib":1166.546875},{"t_s":108,"rss_mib":1166.546875},{"t_s":109,"rss_mib":1166.5546875},{"t_s":111,"rss_mib":1166.5546875},{"t_s":113,"rss_mib":1166.66796875},{"t_s":114,"rss_mib":1166.66796875},{"t_s":116,"rss_mib":1166.6796875},{"t_s":117,"rss_mib":1166.6796875},{"t_s":119,"rss_mib":1166.7109375},{"t_s":120,"rss_mib":1166.7109375},{"t_s":121,"rss_mib":1166.72265625},{"t_s":123,"rss_mib":1166.72265625},{"t_s":125,"rss_mib":1166.76953125},{"t_s":126,"rss_mib":1166.76953125},{"t_s":127,"rss_mib":1166.77734375},{"t_s":129,"rss_mib":1166.77734375},{"t_s":130,"rss_mib":1166.80078125},{"t_s":132,"rss_mib":1166.80078125},{"t_s":134,"rss_mib":1166.83984375},{"t_s":135,"rss_mib":1166.83984375},{"t_s":137,"rss_mib":1166.89453125},{"t_s":138,"rss_mib":1166.89453125},{"t_s":140,"rss_mib":1166.9140625},{"t_s":141,"rss_mib":1166.9140625},{"t_s":144,"rss_mib":1167.09375},{"t_s":144,"rss_mib":1167.09375},{"t_s":146,"rss_mib":1167.1015625},{"t_s":147,"rss_mib":1167.1015625},{"t_s":149,"rss_mib":1167.1171875},{"t_s":150,"rss_mib":1167.12109375},{"t_s":152,"rss_mib":1167.13671875},{"t_s":153,"rss_mib":1167.13671875},{"t_s":154,"rss_mib":1167.14453125},{"t_s":156,"rss_mib":1167.15234375},{"t_s":158,"rss_mib":1167.1640625},{"t_s":159,"rss_mib":1167.1640625},{"t_s":161,"rss_mib":1167.1953125},{"t_s":162,"rss_mib":1167.20703125},{"t_s":162,"rss_mib":1167.21875},{"t_s":165,"rss_mib":1167.21875},{"t_s":167,"rss_mib":1167.24609375},{"t_s":168,"rss_mib":1167.30859375},{"t_s":168,"rss_mib":1163.5234375},{"t_s":171,"rss_mib":1163.5234375},{"t_s":173,"rss_mib":1163.5625},{"t_s":174,"rss_mib":1163.5625},{"t_s":176,"rss_mib":1160.43359375},{"t_s":178,"rss_mib":1160.4375},{"t_s":178,"rss_mib":1159.8828125},{"t_s":180,"rss_mib":1159.8828125},{"t_s":183,"rss_mib":1159.8828125},{"t_s":186,"rss_mib":1159.8828125},{"t_s":189,"rss_mib":1159.8984375},{"t_s":189,"rss_mib":1159.8984375},{"t_s":189,"rss_mib":1157.890625},{"t_s":192,"rss_mib":1157.9296875},{"t_s":192,"rss_mib":1157.97265625},{"t_s":195,"rss_mib":1157.97265625},{"t_s":195,"rss_mib":1157.98046875},{"t_s":198,"rss_mib":1157.98046875},{"t_s":199,"rss_mib":1157.98828125},{"t_s":201,"rss_mib":1157.98828125},{"t_s":203,"rss_mib":1158.15234375},{"t_s":204,"rss_mib":1158.15234375},{"t_s":207,"rss_mib":1158.16796875},{"t_s":207,"rss_mib":1158.16796875},{"t_s":210,"rss_mib":1158.16796875},{"t_s":210,"rss_mib":1158.17578125},{"t_s":213,"rss_mib":1158.17578125},{"t_s":214,"rss_mib":1158.18359375},{"t_s":216,"rss_mib":1158.18359375},{"t_s":217,"rss_mib":1158.19921875},{"t_s":219,"rss_mib":1158.19921875},{"t_s":221,"rss_mib":1158.21875},{"t_s":222,"rss_mib":1158.234375},{"t_s":224,"rss_mib":1158.296875},{"t_s":225,"rss_mib":1158.296875},{"t_s":227,"rss_mib":1158.3203125},{"t_s":228,"rss_mib":1158.3203125},{"t_s":229,"rss_mib":1158.328125},{"t_s":231,"rss_mib":1158.328125},{"t_s":234,"rss_mib":1158.46875},{"t_s":234,"rss_mib":1158.46875},{"t_s":237,"rss_mib":1158.4921875},{"t_s":237,"rss_mib":1158.4921875},{"t_s":238,"rss_mib":1158.50390625},{"t_s":240,"rss_mib":1158.50390625},{"t_s":241,"rss_mib":1158.51171875},{"t_s":243,"rss_mib":1158.51171875},{"t_s":243,"rss_mib":1158.51953125},{"t_s":246,"rss_mib":1158.51953125},{"t_s":247,"rss_mib":1158.52734375},{"t_s":249,"rss_mib":1158.52734375},{"t_s":252,"rss_mib":1158.54296875},{"t_s":252,"rss_mib":1158.54296875},{"t_s":255,"rss_mib":1158.54296875},{"t_s":257,"rss_mib":1158.703125},{"t_s":258,"rss_mib":1158.703125},{"t_s":261,"rss_mib":1158.703125},{"t_s":263,"rss_mib":1158.875},{"t_s":264,"rss_mib":1158.875},{"t_s":266,"rss_mib":1158.8828125},{"t_s":267,"rss_mib":1158.8828125},{"t_s":270,"rss_mib":1158.890625},{"t_s":273,"rss_mib":1158.890625},{"t_s":274,"rss_mib":1158.8984375},{"t_s":276,"rss_mib":1158.8984375},{"t_s":277,"rss_mib":1158.91015625},{"t_s":279,"rss_mib":1158.91015625},{"t_s":282,"rss_mib":1158.91796875},{"t_s":282,"rss_mib":1158.91796875},{"t_s":285,"rss_mib":1158.9453125},{"t_s":285,"rss_mib":1158.9453125},{"t_s":287,"rss_mib":1158.9609375},{"t_s":288,"rss_mib":1158.98828125},{"t_s":290,"rss_mib":1159.01953125},{"t_s":291,"rss_mib":1159.01953125},{"t_s":293,"rss_mib":1159.13671875},{"t_s":294,"rss_mib":1159.13671875},{"t_s":297,"rss_mib":1159.13671875},{"t_s":299,"rss_mib":1159.1484375},{"t_s":300,"rss_mib":1159.1484375},{"t_s":303,"rss_mib":1159.1484375},{"t_s":306,"rss_mib":1159.1484375},{"t_s":308,"rss_mib":1159.52734375},{"t_s":309,"rss_mib":1159.52734375},{"t_s":312,"rss_mib":1159.52734375},{"t_s":315,"rss_mib":1159.52734375},{"t_s":317,"rss_mib":1159.5546875},{"t_s":318,"rss_mib":1159.5546875},{"t_s":319,"rss_mib":1159.59375},{"t_s":321,"rss_mib":1159.59375},{"t_s":323,"rss_mib":1159.671875},{"t_s":324,"rss_mib":1159.671875},{"t_s":326,"rss_mib":1159.6796875},{"t_s":327,"rss_mib":1159.6796875},{"t_s":329,"rss_mib":1159.6953125},{"t_s":330,"rss_mib":1159.6953125},{"t_s":331,"rss_mib":1159.703125},{"t_s":333,"rss_mib":1159.703125},{"t_s":334,"rss_mib":1159.7109375},{"t_s":336,"rss_mib":1159.7109375},{"t_s":338,"rss_mib":1159.75},{"t_s":339,"rss_mib":1159.75},{"t_s":341,"rss_mib":1159.765625},{"t_s":342,"rss_mib":1159.765625},{"t_s":343,"rss_mib":1159.7734375},{"t_s":345,"rss_mib":1159.7734375},{"t_s":347,"rss_mib":1159.7890625},{"t_s":348,"rss_mib":1159.7890625},{"t_s":351,"rss_mib":1159.80859375},{"t_s":351,"rss_mib":1159.80859375},{"t_s":354,"rss_mib":1159.91015625},{"t_s":354,"rss_mib":1159.91015625},{"t_s":356,"rss_mib":1159.92578125},{"t_s":357,"rss_mib":1159.92578125},{"t_s":360,"rss_mib":1159.94921875},{"t_s":360,"rss_mib":1159.94921875}],"idle_rss_series":[{"t_s":0,"rss_mib":1147.82421875},{"t_s":0,"rss_mib":1147.90625},{"t_s":0,"rss_mib":1147.9140625},{"t_s":0,"rss_mib":1147.921875},{"t_s":0,"rss_mib":1147.921875},{"t_s":1,"rss_mib":1147.921875},{"t_s":1,"rss_mib":1148.01953125},{"t_s":2,"rss_mib":1148.02734375},{"t_s":2,"rss_mib":1148.02734375},{"t_s":3,"rss_mib":1148.02734375},{"t_s":3,"rss_mib":1148.02734375},{"t_s":4,"rss_mib":1148.03515625},{"t_s":4,"rss_mib":1148.04296875},{"t_s":4,"rss_mib":1148.04296875},{"t_s":5,"rss_mib":1148.04296875},{"t_s":5,"rss_mib":1148.04296875},{"t_s":6,"rss_mib":1148.0546875},{"t_s":6,"rss_mib":1148.0546875},{"t_s":6,"rss_mib":1148.0625},{"t_s":6,"rss_mib":1148.0625},{"t_s":7,"rss_mib":1148.0703125},{"t_s":7,"rss_mib":1148.0703125},{"t_s":7,"rss_mib":1148.0703125},{"t_s":8,"rss_mib":1148.0703125},{"t_s":8,"rss_mib":1148.078125},{"t_s":8,"rss_mib":1148.0859375},{"t_s":9,"rss_mib":1148.0859375},{"t_s":9,"rss_mib":1148.09375},{"t_s":10,"rss_mib":1148.09375},{"t_s":10,"rss_mib":1148.09375},{"t_s":10,"rss_mib":1148.1015625},{"t_s":10,"rss_mib":1148.1015625},{"t_s":11,"rss_mib":1148.1015625},{"t_s":11,"rss_mib":1148.109375},{"t_s":11,"rss_mib":1148.109375},{"t_s":12,"rss_mib":1148.109375},{"t_s":12,"rss_mib":1148.1171875},{"t_s":13,"rss_mib":1148.1171875},{"t_s":13,"rss_mib":1148.125},{"t_s":13,"rss_mib":1148.125},{"t_s":13,"rss_mib":1148.1328125},{"t_s":13,"rss_mib":1148.1328125},{"t_s":14,"rss_mib":1148.1328125},{"t_s":14,"rss_mib":1148.1328125},{"t_s":15,"rss_mib":1148.140625},{"t_s":15,"rss_mib":1148.140625},{"t_s":15,"rss_mib":1148.140625},{"t_s":16,"rss_mib":1148.15625},{"t_s":16,"rss_mib":1148.15625},{"t_s":16,"rss_mib":1148.15625},{"t_s":17,"rss_mib":1148.16796875},{"t_s":17,"rss_mib":1148.16796875},{"t_s":17,"rss_mib":1148.17578125},{"t_s":17,"rss_mib":1148.18359375},{"t_s":18,"rss_mib":1148.18359375},{"t_s":18,"rss_mib":1148.18359375},{"t_s":19,"rss_mib":1148.19140625},{"t_s":19,"rss_mib":1148.19140625},{"t_s":20,"rss_mib":1148.19140625},{"t_s":20,"rss_mib":1148.2109375},{"t_s":20,"rss_mib":1148.2109375},{"t_s":20,"rss_mib":1148.2109375},{"t_s":21,"rss_mib":1148.2109375},{"t_s":21,"rss_mib":1148.21484375},{"t_s":21,"rss_mib":1148.21484375},{"t_s":22,"rss_mib":1148.2265625},{"t_s":22,"rss_mib":1148.2265625},{"t_s":22,"rss_mib":1148.234375},{"t_s":23,"rss_mib":1148.2421875},{"t_s":23,"rss_mib":1148.33203125},{"t_s":23,"rss_mib":1148.33203125},{"t_s":24,"rss_mib":1148.33984375},{"t_s":24,"rss_mib":1148.33984375},{"t_s":24,"rss_mib":1148.35546875},{"t_s":24,"rss_mib":1148.37109375},{"t_s":25,"rss_mib":1148.37890625},{"t_s":25,"rss_mib":1148.37890625},{"t_s":26,"rss_mib":1148.38671875},{"t_s":26,"rss_mib":1148.390625},{"t_s":26,"rss_mib":1148.390625},{"t_s":26,"rss_mib":1148.3984375},{"t_s":26,"rss_mib":1148.3984375},{"t_s":27,"rss_mib":1148.3984375},{"t_s":27,"rss_mib":1148.4140625},{"t_s":28,"rss_mib":1148.4140625},{"t_s":29,"rss_mib":1148.4140625},{"t_s":29,"rss_mib":1148.421875},{"t_s":30,"rss_mib":1148.421875},{"t_s":30,"rss_mib":1148.4296875},{"t_s":30,"rss_mib":1148.4375},{"t_s":30,"rss_mib":1148.4375},{"t_s":31,"rss_mib":1148.4375},{"t_s":31,"rss_mib":1148.4453125},{"t_s":31,"rss_mib":1148.4453125},{"t_s":32,"rss_mib":1148.4453125},{"t_s":33,"rss_mib":1148.4453125},{"t_s":33,"rss_mib":1148.4609375},{"t_s":33,"rss_mib":1148.4609375},{"t_s":33,"rss_mib":1148.4609375},{"t_s":34,"rss_mib":1148.46875},{"t_s":34,"rss_mib":1148.46875},{"t_s":35,"rss_mib":1148.4765625},{"t_s":35,"rss_mib":1148.4765625},{"t_s":35,"rss_mib":1148.484375},{"t_s":36,"rss_mib":1148.484375},{"t_s":36,"rss_mib":1148.484375},{"t_s":36,"rss_mib":1148.4921875},{"t_s":36,"rss_mib":1148.4921875},{"t_s":37,"rss_mib":1148.4921875},{"t_s":37,"rss_mib":1148.5},{"t_s":37,"rss_mib":1148.5},{"t_s":38,"rss_mib":1148.5},{"t_s":38,"rss_mib":1148.5078125},{"t_s":39,"rss_mib":1148.5078125},{"t_s":39,"rss_mib":1148.51953125},{"t_s":39,"rss_mib":1148.51953125},{"t_s":40,"rss_mib":1148.51953125},{"t_s":40,"rss_mib":1148.52734375},{"t_s":40,"rss_mib":1148.52734375},{"t_s":40,"rss_mib":1148.52734375},{"t_s":41,"rss_mib":1148.53125},{"t_s":41,"rss_mib":1148.53125},{"t_s":41,"rss_mib":1148.5390625},{"t_s":42,"rss_mib":1148.5390625},{"t_s":42,"rss_mib":1148.5546875},{"t_s":42,"rss_mib":1148.5546875},{"t_s":43,"rss_mib":1148.5546875},{"t_s":43,"rss_mib":1148.5546875},{"t_s":43,"rss_mib":1148.5546875},{"t_s":44,"rss_mib":1148.5546875},{"t_s":44,"rss_mib":1148.5625},{"t_s":44,"rss_mib":1148.5625},{"t_s":45,"rss_mib":1148.578125},{"t_s":45,"rss_mib":1148.578125},{"t_s":46,"rss_mib":1148.5859375},{"t_s":46,"rss_mib":1148.59375},{"t_s":46,"rss_mib":1148.59375},{"t_s":46,"rss_mib":1148.59375},{"t_s":47,"rss_mib":1148.59765625},{"t_s":47,"rss_mib":1148.59765625},{"t_s":47,"rss_mib":1148.60546875},{"t_s":47,"rss_mib":1148.60546875},{"t_s":48,"rss_mib":1148.61328125},{"t_s":48,"rss_mib":1148.61328125},{"t_s":49,"rss_mib":1148.61328125},{"t_s":49,"rss_mib":1148.62109375},{"t_s":49,"rss_mib":1148.62109375},{"t_s":49,"rss_mib":1148.62890625},{"t_s":50,"rss_mib":1148.62890625},{"t_s":50,"rss_mib":1148.62890625},{"t_s":51,"rss_mib":1148.63671875},{"t_s":51,"rss_mib":1148.63671875},{"t_s":51,"rss_mib":1148.65625},{"t_s":52,"rss_mib":1148.65625},{"t_s":52,"rss_mib":1148.66015625},{"t_s":52,"rss_mib":1148.66015625},{"t_s":52,"rss_mib":1148.66796875},{"t_s":53,"rss_mib":1148.66796875},{"t_s":53,"rss_mib":1148.67578125},{"t_s":53,"rss_mib":1148.78125},{"t_s":53,"rss_mib":1148.7890625},{"t_s":54,"rss_mib":1148.7890625},{"t_s":54,"rss_mib":1148.8046875},{"t_s":55,"rss_mib":1148.8046875},{"t_s":55,"rss_mib":1148.81640625},{"t_s":55,"rss_mib":1148.81640625},{"t_s":55,"rss_mib":1148.82421875},{"t_s":56,"rss_mib":1148.82421875},{"t_s":56,"rss_mib":1148.82421875},{"t_s":56,"rss_mib":1148.83203125},{"t_s":56,"rss_mib":1148.83203125},{"t_s":57,"rss_mib":1148.84765625},{"t_s":57,"rss_mib":1148.84765625},{"t_s":58,"rss_mib":1148.84765625},{"t_s":58,"rss_mib":1148.86328125},{"t_s":58,"rss_mib":1148.86328125},{"t_s":59,"rss_mib":1148.86328125},{"t_s":59,"rss_mib":1148.86328125},{"t_s":59,"rss_mib":1148.87890625},{"t_s":59,"rss_mib":1148.87890625}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.010302158,"cost":6.035073029,"memory":434.937408147,"streaming":2.749736265,"streams_sustained":681.178598185,"throughput":331.270885977},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"perf.frontier.1ms.concurrency":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.p99_us":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.rps":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.samples":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"stream.added_gap_p99_us":{"reason":"below_resolution","detail":"the gateway's own inter-frame gap at this percentile (21173us) came in under the mock's (21174us), which a proxy cannot really do - the added gap here is below what this rig can resolve against the mock's 20ms pacing"}}},"bedrock":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/model/claude-3-5-sonnet-20241022/converse","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"openai-responses":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1/responses","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"anthropic":{"served":true,"reason":null,"status":"","path":"/v1/messages","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":680,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":822,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":850,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":674.8307904095445,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1481.852953676158,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.9771911881943663,"certified":true,"suppressed":false},"cost_window_ok":{"value":34868,"certified":true,"suppressed":false},"cost_window_rps":{"value":5810,"certified":true,"suppressed":false},"cost_threads":{"value":144,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.0020075714121830907,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":1352,"certified":true,"suppressed":false},"concurrency":1,"p99_us":957,"first_disqualified_conc":2,"lower_bound":false,"samples":8115},{"bound_ms":5,"rps":{"value":5967,"certified":true,"suppressed":false},"concurrency":16,"p99_us":4103,"first_disqualified_conc":32,"lower_bound":false,"samples":35809},{"bound_ms":10,"rps":{"value":6123,"certified":true,"suppressed":false},"concurrency":32,"p99_us":9073,"first_disqualified_conc":64,"lower_bound":false,"samples":36743},{"bound_ms":50,"rps":{"value":6224,"certified":true,"suppressed":false},"concurrency":128,"p99_us":27064,"first_disqualified_conc":256,"lower_bound":false,"samples":37352},{"bound_ms":100,"rps":{"value":6224,"certified":true,"suppressed":false},"concurrency":128,"p99_us":27064,"first_disqualified_conc":512,"lower_bound":false,"samples":37352},{"bound_ms":null,"rps":{"value":6224,"certified":true,"suppressed":false},"concurrency":128,"p99_us":27064,"first_disqualified_conc":null,"lower_bound":false,"samples":37352}],"sweep":[{"conc":1,"ok":8109,"rps":1351,"p99_us":961,"fail":0},{"conc":1,"ok":8119,"rps":1352,"p99_us":959,"fail":0},{"conc":1,"ok":8115,"rps":1352,"p99_us":957,"fail":0},{"conc":2,"ok":8860,"rps":1476,"p99_us":1672,"fail":0},{"conc":2,"ok":8869,"rps":1477,"p99_us":1668,"fail":0},{"conc":2,"ok":8860,"rps":1476,"p99_us":1671,"fail":0},{"conc":4,"ok":24978,"rps":4162,"p99_us":1665,"fail":0},{"conc":4,"ok":16963,"rps":2826,"p99_us":2519,"fail":0},{"conc":4,"ok":25097,"rps":4182,"p99_us":1660,"fail":0},{"conc":8,"ok":34418,"rps":5735,"p99_us":2419,"fail":0},{"conc":8,"ok":33851,"rps":5640,"p99_us":3224,"fail":0},{"conc":8,"ok":26552,"rps":4424,"p99_us":2564,"fail":0},{"conc":16,"ok":35812,"rps":5968,"p99_us":5520,"fail":0},{"conc":16,"ok":35685,"rps":5946,"p99_us":4122,"fail":0},{"conc":16,"ok":35809,"rps":5967,"p99_us":4103,"fail":0},{"conc":32,"ok":36678,"rps":6112,"p99_us":8739,"fail":0},{"conc":32,"ok":36743,"rps":6123,"p99_us":9073,"fail":0},{"conc":32,"ok":36716,"rps":6118,"p99_us":8032,"fail":0},{"conc":64,"ok":37255,"rps":6208,"p99_us":17935,"fail":0},{"conc":64,"ok":37288,"rps":6213,"p99_us":16449,"fail":0},{"conc":64,"ok":37289,"rps":6213,"p99_us":17612,"fail":0},{"conc":128,"ok":37191,"rps":6196,"p99_us":28471,"fail":0},{"conc":128,"ok":37287,"rps":6213,"p99_us":29532,"fail":0},{"conc":128,"ok":37352,"rps":6224,"p99_us":27064,"fail":0},{"conc":256,"ok":35826,"rps":5969,"p99_us":59099,"fail":0},{"conc":256,"ok":35970,"rps":5994,"p99_us":58298,"fail":0},{"conc":256,"ok":35700,"rps":5949,"p99_us":63480,"fail":0},{"conc":512,"ok":34683,"rps":5779,"p99_us":112828,"fail":0},{"conc":512,"ok":34696,"rps":5781,"p99_us":122441,"fail":0},{"conc":512,"ok":34647,"rps":5773,"p99_us":119043,"fail":0},{"conc":1024,"ok":34759,"rps":5791,"p99_us":217281,"fail":0},{"conc":1024,"ok":34634,"rps":5770,"p99_us":306642,"fail":0},{"conc":1024,"ok":34604,"rps":5766,"p99_us":227056,"fail":0},{"conc":2048,"ok":34850,"rps":5807,"p99_us":755241,"fail":0},{"conc":2048,"ok":34726,"rps":5786,"p99_us":711286,"fail":0},{"conc":2048,"ok":34864,"rps":5809,"p99_us":781351,"fail":0},{"conc":4096,"ok":34796,"rps":5798,"p99_us":778016,"fail":0},{"conc":4096,"ok":34921,"rps":5818,"p99_us":718696,"fail":0},{"conc":4096,"ok":34882,"rps":5812,"p99_us":763325,"fail":0},{"conc":8192,"ok":34862,"rps":5809,"p99_us":533952,"fail":0},{"conc":8192,"ok":34881,"rps":5813,"p99_us":809146,"fail":0},{"conc":8192,"ok":34876,"rps":5811,"p99_us":803706,"fail":0},{"conc":16384,"ok":34948,"rps":5824,"p99_us":905333,"fail":0},{"conc":16384,"ok":34955,"rps":5824,"p99_us":571248,"fail":0},{"conc":16384,"ok":34842,"rps":5805,"p99_us":1122173,"fail":0},{"conc":32768,"ok":26716,"rps":4451,"p99_us":3061286,"fail":0},{"conc":32768,"ok":26320,"rps":4386,"p99_us":2628440,"fail":0},{"conc":32768,"ok":31814,"rps":5300,"p99_us":2612280,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1352,"certified":true,"suppressed":false},"p99_us":959,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":1476,"certified":true,"suppressed":false},"p99_us":1671,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":4162,"certified":true,"suppressed":false},"p99_us":1665,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":5640,"certified":true,"suppressed":false},"p99_us":2564,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":5967,"certified":true,"suppressed":false},"p99_us":4122,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":6118,"certified":true,"suppressed":false},"p99_us":8739,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":6213,"certified":true,"suppressed":false},"p99_us":17612,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":6213,"certified":true,"suppressed":false},"p99_us":28471,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":5969,"certified":true,"suppressed":false},"p99_us":59099,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":5779,"certified":true,"suppressed":false},"p99_us":119043,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":5770,"certified":true,"suppressed":false},"p99_us":227056,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":5807,"certified":true,"suppressed":false},"p99_us":755241,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":5812,"certified":true,"suppressed":false},"p99_us":763325,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":5811,"certified":true,"suppressed":false},"p99_us":803706,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":5824,"certified":true,"suppressed":false},"p99_us":905333,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":4451,"certified":true,"suppressed":false},"p99_us":2628440,"windows":3,"clean_windows":3}],"reverify_note":"anthropic>anthropic is a same-dialect cell: there is no translation to prove, so a request-shape check upstream cannot distinguish a translating gateway from a forwarding one","c1_note":"the c=1 percentiles are taken over 8481 successful gateway round trip(s) and 262245 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":8481,"direct_c1_samples":262245},"stream":{"stream_served":true,"stream_error":null,"reason":null,"stream_c1_note":"the c=1 streaming legs read 64 frame(s) through the gateway and 64 direct to the mock, out of a 64 frame budget. The added-GAP figures are percentiles over the intervals inside those two streams. The added-TTFT figures are percentiles over separate probes - 100 through the gateway and 100 direct, out of 100 attempted per leg - so they carry that weight and no more","added_ttft_p50_us":{"value":822,"certified":true,"suppressed":false},"added_ttft_p99_us":{"value":1012,"certified":true,"suppressed":false},"added_gap_p50_us":{"value":0,"certified":true,"suppressed":false},"added_gap_p99_us":{"value":35,"certified":true,"suppressed":false},"streams_sustained_fps":{"value":52509.65359935977,"certified":true,"suppressed":false,"headroom":0.06307020958070865,"rig_ceiling":832558.7301587302},"streams_sustained":{"value":16391,"certified":true,"suppressed":false,"headroom":0.06307020958070865,"rig_ceiling":832558.7301587302},"ttft_gw_samples":100,"ttft_direct_samples":100,"gap_gw_samples":63,"gap_direct_samples":63},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1153.921875,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1161.83984375,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1162.21484375,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1171.8125,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1272.52734375,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.275204255009843,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":0.5817656189913647,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1154.58203125},{"t_s":2,"rss_mib":1171.0078125},{"t_s":3,"rss_mib":1171.0390625},{"t_s":5,"rss_mib":1171.12109375},{"t_s":6,"rss_mib":1171.12109375},{"t_s":7,"rss_mib":1171.3046875},{"t_s":9,"rss_mib":1171.3046875},{"t_s":10,"rss_mib":1171.3125},{"t_s":12,"rss_mib":1171.3125},{"t_s":14,"rss_mib":1171.32421875},{"t_s":15,"rss_mib":1171.32421875},{"t_s":15,"rss_mib":1171.3359375},{"t_s":18,"rss_mib":1171.3359375},{"t_s":19,"rss_mib":1171.4765625},{"t_s":21,"rss_mib":1171.4765625},{"t_s":22,"rss_mib":1171.5},{"t_s":24,"rss_mib":1171.5},{"t_s":24,"rss_mib":1171.5390625},{"t_s":27,"rss_mib":1171.5390625},{"t_s":30,"rss_mib":1171.5390625},{"t_s":31,"rss_mib":1171.6953125},{"t_s":33,"rss_mib":1171.6953125},{"t_s":35,"rss_mib":1171.7265625},{"t_s":36,"rss_mib":1171.7265625},{"t_s":39,"rss_mib":1171.7265625},{"t_s":42,"rss_mib":1171.7265625},{"t_s":44,"rss_mib":1171.734375},{"t_s":45,"rss_mib":1171.734375},{"t_s":47,"rss_mib":1171.73828125},{"t_s":48,"rss_mib":1171.73828125},{"t_s":49,"rss_mib":1171.7890625},{"t_s":51,"rss_mib":1171.7890625},{"t_s":52,"rss_mib":1171.80078125},{"t_s":54,"rss_mib":1171.8125},{"t_s":55,"rss_mib":1168.8203125},{"t_s":59,"rss_mib":1168.88671875},{"t_s":59,"rss_mib":1166.58984375},{"t_s":60,"rss_mib":1166.703125},{"t_s":62,"rss_mib":1166.71875},{"t_s":63,"rss_mib":1166.7265625},{"t_s":66,"rss_mib":1166.7265625},{"t_s":67,"rss_mib":1166.73046875},{"t_s":69,"rss_mib":1166.73046875},{"t_s":72,"rss_mib":1166.734375},{"t_s":72,"rss_mib":1166.7578125},{"t_s":75,"rss_mib":1166.7578125},{"t_s":77,"rss_mib":1166.765625},{"t_s":78,"rss_mib":1166.765625},{"t_s":79,"rss_mib":1166.85546875},{"t_s":81,"rss_mib":1166.85546875},{"t_s":82,"rss_mib":1166.8671875},{"t_s":84,"rss_mib":1166.8671875},{"t_s":84,"rss_mib":1166.875},{"t_s":87,"rss_mib":1166.875},{"t_s":89,"rss_mib":1166.890625},{"t_s":90,"rss_mib":1166.890625},{"t_s":93,"rss_mib":1166.890625},{"t_s":96,"rss_mib":1166.90234375},{"t_s":97,"rss_mib":1166.9140625},{"t_s":99,"rss_mib":1166.9140625},{"t_s":102,"rss_mib":1166.9140625},{"t_s":103,"rss_mib":1166.92578125},{"t_s":105,"rss_mib":1166.92578125},{"t_s":106,"rss_mib":1166.9296875},{"t_s":108,"rss_mib":1166.9296875},{"t_s":109,"rss_mib":1167.015625},{"t_s":111,"rss_mib":1167.015625},{"t_s":114,"rss_mib":1167.03515625},{"t_s":114,"rss_mib":1167.03515625},{"t_s":116,"rss_mib":1167.04296875},{"t_s":117,"rss_mib":1167.04296875},{"t_s":119,"rss_mib":1167.05859375},{"t_s":120,"rss_mib":1167.05859375},{"t_s":122,"rss_mib":1167.20703125},{"t_s":123,"rss_mib":1167.20703125},{"t_s":126,"rss_mib":1167.2109375},{"t_s":126,"rss_mib":1167.2109375},{"t_s":129,"rss_mib":1167.23828125},{"t_s":129,"rss_mib":1167.23828125},{"t_s":132,"rss_mib":1167.265625},{"t_s":132,"rss_mib":1167.265625},{"t_s":134,"rss_mib":1167.27734375},{"t_s":135,"rss_mib":1167.27734375},{"t_s":136,"rss_mib":1167.28515625},{"t_s":138,"rss_mib":1167.28515625},{"t_s":141,"rss_mib":1167.28515625},{"t_s":144,"rss_mib":1167.28515625},{"t_s":147,"rss_mib":1167.28515625},{"t_s":147,"rss_mib":1167.2890625},{"t_s":150,"rss_mib":1167.2890625},{"t_s":152,"rss_mib":1167.296875},{"t_s":153,"rss_mib":1167.296875},{"t_s":156,"rss_mib":1167.296875},{"t_s":158,"rss_mib":1167.31640625},{"t_s":159,"rss_mib":1167.31640625},{"t_s":162,"rss_mib":1167.3203125},{"t_s":162,"rss_mib":1167.3203125},{"t_s":165,"rss_mib":1167.3203125},{"t_s":168,"rss_mib":1167.3359375},{"t_s":168,"rss_mib":1167.3359375},{"t_s":169,"rss_mib":1167.40234375},{"t_s":171,"rss_mib":1167.40234375},{"t_s":174,"rss_mib":1167.40234375},{"t_s":175,"rss_mib":1163.01171875},{"t_s":178,"rss_mib":1163.0234375},{"t_s":179,"rss_mib":1161.328125},{"t_s":180,"rss_mib":1161.328125},{"t_s":182,"rss_mib":1161.34765625},{"t_s":184,"rss_mib":1161.359375},{"t_s":184,"rss_mib":1161.328125},{"t_s":186,"rss_mib":1161.33203125},{"t_s":188,"rss_mib":1161.36328125},{"t_s":189,"rss_mib":1161.36328125},{"t_s":192,"rss_mib":1161.36328125},{"t_s":192,"rss_mib":1161.375},{"t_s":195,"rss_mib":1161.375},{"t_s":197,"rss_mib":1161.3828125},{"t_s":198,"rss_mib":1161.4375},{"t_s":200,"rss_mib":1161.453125},{"t_s":201,"rss_mib":1161.453125},{"t_s":204,"rss_mib":1161.453125},{"t_s":204,"rss_mib":1161.46484375},{"t_s":207,"rss_mib":1161.46484375},{"t_s":208,"rss_mib":1161.47265625},{"t_s":210,"rss_mib":1161.47265625},{"t_s":213,"rss_mib":1161.47265625},{"t_s":214,"rss_mib":1161.48046875},{"t_s":216,"rss_mib":1161.48046875},{"t_s":217,"rss_mib":1161.484375},{"t_s":219,"rss_mib":1161.484375},{"t_s":222,"rss_mib":1161.484375},{"t_s":222,"rss_mib":1161.51171875},{"t_s":225,"rss_mib":1161.51171875},{"t_s":227,"rss_mib":1161.52734375},{"t_s":228,"rss_mib":1161.52734375},{"t_s":231,"rss_mib":1161.61328125},{"t_s":231,"rss_mib":1161.67578125},{"t_s":232,"rss_mib":1161.6875},{"t_s":234,"rss_mib":1161.6875},{"t_s":237,"rss_mib":1161.6875},{"t_s":240,"rss_mib":1161.69140625},{"t_s":240,"rss_mib":1161.69140625},{"t_s":240,"rss_mib":1163.10546875},{"t_s":243,"rss_mib":1161.75},{"t_s":246,"rss_mib":1161.75},{"t_s":247,"rss_mib":1161.7890625},{"t_s":249,"rss_mib":1161.7890625},{"t_s":252,"rss_mib":1161.7890625},{"t_s":254,"rss_mib":1161.84765625},{"t_s":255,"rss_mib":1161.84765625},{"t_s":258,"rss_mib":1161.859375},{"t_s":258,"rss_mib":1161.859375},{"t_s":259,"rss_mib":1161.86328125},{"t_s":261,"rss_mib":1161.86328125},{"t_s":264,"rss_mib":1161.86328125},{"t_s":267,"rss_mib":1161.86328125},{"t_s":270,"rss_mib":1161.86328125},{"t_s":270,"rss_mib":1161.8671875},{"t_s":273,"rss_mib":1161.8671875},{"t_s":276,"rss_mib":1161.8671875},{"t_s":278,"rss_mib":1161.87890625},{"t_s":279,"rss_mib":1161.87890625},{"t_s":282,"rss_mib":1161.87890625},{"t_s":285,"rss_mib":1161.87890625},{"t_s":288,"rss_mib":1161.90234375},{"t_s":288,"rss_mib":1161.91015625},{"t_s":289,"rss_mib":1162.01171875},{"t_s":291,"rss_mib":1162.01171875},{"t_s":293,"rss_mib":1162.01953125},{"t_s":294,"rss_mib":1162.01953125},{"t_s":296,"rss_mib":1162.09765625},{"t_s":297,"rss_mib":1162.09765625},{"t_s":299,"rss_mib":1162.109375},{"t_s":300,"rss_mib":1162.109375},{"t_s":300,"rss_mib":1162.125},{"t_s":303,"rss_mib":1162.125},{"t_s":304,"rss_mib":1162.109375},{"t_s":306,"rss_mib":1162.1171875},{"t_s":309,"rss_mib":1162.1171875},{"t_s":310,"rss_mib":1162.12109375},{"t_s":312,"rss_mib":1162.12109375},{"t_s":314,"rss_mib":1162.18359375},{"t_s":315,"rss_mib":1162.18359375},{"t_s":316,"rss_mib":1162.19140625},{"t_s":318,"rss_mib":1162.19140625},{"t_s":321,"rss_mib":1162.19140625},{"t_s":324,"rss_mib":1162.19140625},{"t_s":325,"rss_mib":1162.19921875},{"t_s":327,"rss_mib":1162.19921875},{"t_s":330,"rss_mib":1162.19921875},{"t_s":330,"rss_mib":1162.203125},{"t_s":333,"rss_mib":1162.203125},{"t_s":336,"rss_mib":1162.203125},{"t_s":337,"rss_mib":1162.20703125},{"t_s":339,"rss_mib":1162.20703125},{"t_s":340,"rss_mib":1162.2109375},{"t_s":342,"rss_mib":1162.2109375},{"t_s":343,"rss_mib":1162.21484375},{"t_s":345,"rss_mib":1162.21484375},{"t_s":348,"rss_mib":1162.22265625},{"t_s":349,"rss_mib":1162.2890625},{"t_s":351,"rss_mib":1162.2890625},{"t_s":354,"rss_mib":1162.2890625},{"t_s":355,"rss_mib":1162.29296875},{"t_s":357,"rss_mib":1162.29296875},{"t_s":357,"rss_mib":1162.3046875},{"t_s":360,"rss_mib":1162.3046875}],"idle_rss_series":[{"t_s":0,"rss_mib":1153.44921875},{"t_s":0,"rss_mib":1153.52734375},{"t_s":0,"rss_mib":1153.52734375},{"t_s":0,"rss_mib":1153.578125},{"t_s":0,"rss_mib":1153.578125},{"t_s":1,"rss_mib":1153.578125},{"t_s":1,"rss_mib":1153.578125},{"t_s":2,"rss_mib":1153.578125},{"t_s":3,"rss_mib":1153.578125},{"t_s":3,"rss_mib":1153.578125},{"t_s":3,"rss_mib":1153.578125},{"t_s":4,"rss_mib":1153.578125},{"t_s":4,"rss_mib":1153.77734375},{"t_s":5,"rss_mib":1153.77734375},{"t_s":5,"rss_mib":1153.77734375},{"t_s":6,"rss_mib":1153.77734375},{"t_s":6,"rss_mib":1153.77734375},{"t_s":7,"rss_mib":1153.77734375},{"t_s":7,"rss_mib":1153.78125},{"t_s":7,"rss_mib":1153.78125},{"t_s":8,"rss_mib":1153.78125},{"t_s":8,"rss_mib":1153.78125},{"t_s":9,"rss_mib":1153.78125},{"t_s":9,"rss_mib":1153.78125},{"t_s":10,"rss_mib":1153.78515625},{"t_s":10,"rss_mib":1153.78515625},{"t_s":10,"rss_mib":1153.78515625},{"t_s":11,"rss_mib":1153.78515625},{"t_s":11,"rss_mib":1153.78515625},{"t_s":12,"rss_mib":1153.7890625},{"t_s":12,"rss_mib":1153.7890625},{"t_s":12,"rss_mib":1153.7890625},{"t_s":13,"rss_mib":1153.7890625},{"t_s":13,"rss_mib":1153.7890625},{"t_s":14,"rss_mib":1153.7890625},{"t_s":14,"rss_mib":1153.7890625},{"t_s":15,"rss_mib":1153.7890625},{"t_s":15,"rss_mib":1153.7890625},{"t_s":16,"rss_mib":1153.7890625},{"t_s":16,"rss_mib":1153.7890625},{"t_s":17,"rss_mib":1153.7890625},{"t_s":17,"rss_mib":1153.796875},{"t_s":18,"rss_mib":1153.796875},{"t_s":18,"rss_mib":1153.796875},{"t_s":19,"rss_mib":1153.796875},{"t_s":19,"rss_mib":1153.796875},{"t_s":20,"rss_mib":1153.796875},{"t_s":20,"rss_mib":1153.796875},{"t_s":21,"rss_mib":1153.796875},{"t_s":22,"rss_mib":1153.796875},{"t_s":22,"rss_mib":1153.80859375},{"t_s":22,"rss_mib":1153.80859375},{"t_s":22,"rss_mib":1153.80859375},{"t_s":23,"rss_mib":1153.81640625},{"t_s":23,"rss_mib":1153.81640625},{"t_s":23,"rss_mib":1153.81640625},{"t_s":24,"rss_mib":1153.81640625},{"t_s":24,"rss_mib":1153.82421875},{"t_s":25,"rss_mib":1153.83984375},{"t_s":25,"rss_mib":1153.85546875},{"t_s":25,"rss_mib":1153.85546875},{"t_s":25,"rss_mib":1153.86328125},{"t_s":25,"rss_mib":1153.86328125},{"t_s":26,"rss_mib":1153.8671875},{"t_s":26,"rss_mib":1153.875},{"t_s":26,"rss_mib":1153.875},{"t_s":27,"rss_mib":1153.88671875},{"t_s":27,"rss_mib":1153.88671875},{"t_s":27,"rss_mib":1153.8984375},{"t_s":28,"rss_mib":1153.90625},{"t_s":28,"rss_mib":1153.90625},{"t_s":28,"rss_mib":1153.90625},{"t_s":29,"rss_mib":1153.9140625},{"t_s":29,"rss_mib":1153.921875},{"t_s":29,"rss_mib":1153.921875},{"t_s":30,"rss_mib":1153.921875},{"t_s":31,"rss_mib":1153.9296875},{"t_s":31,"rss_mib":1153.94140625},{"t_s":31,"rss_mib":1153.94140625},{"t_s":31,"rss_mib":1153.94140625},{"t_s":32,"rss_mib":1153.94140625},{"t_s":32,"rss_mib":1153.94921875},{"t_s":32,"rss_mib":1153.94921875},{"t_s":33,"rss_mib":1153.94921875},{"t_s":33,"rss_mib":1153.94921875},{"t_s":34,"rss_mib":1153.94921875},{"t_s":34,"rss_mib":1153.95703125},{"t_s":34,"rss_mib":1153.95703125},{"t_s":35,"rss_mib":1153.95703125},{"t_s":35,"rss_mib":1153.95703125},{"t_s":36,"rss_mib":1153.95703125},{"t_s":36,"rss_mib":1153.96484375},{"t_s":37,"rss_mib":1153.96484375},{"t_s":37,"rss_mib":1153.96484375},{"t_s":38,"rss_mib":1153.96484375},{"t_s":38,"rss_mib":1153.97265625},{"t_s":38,"rss_mib":1153.97265625},{"t_s":39,"rss_mib":1153.98046875},{"t_s":39,"rss_mib":1153.98046875},{"t_s":39,"rss_mib":1153.98046875},{"t_s":40,"rss_mib":1153.98046875},{"t_s":41,"rss_mib":1153.98046875},{"t_s":41,"rss_mib":1153.98046875},{"t_s":41,"rss_mib":1153.98046875},{"t_s":42,"rss_mib":1153.98046875},{"t_s":42,"rss_mib":1153.98828125},{"t_s":42,"rss_mib":1153.98828125},{"t_s":43,"rss_mib":1153.9921875},{"t_s":43,"rss_mib":1153.9921875},{"t_s":44,"rss_mib":1153.9921875},{"t_s":44,"rss_mib":1153.9921875},{"t_s":44,"rss_mib":1153.9921875},{"t_s":45,"rss_mib":1153.9921875},{"t_s":45,"rss_mib":1153.9921875},{"t_s":46,"rss_mib":1153.9921875},{"t_s":46,"rss_mib":1154},{"t_s":47,"rss_mib":1154},{"t_s":47,"rss_mib":1154},{"t_s":47,"rss_mib":1154.0078125},{"t_s":47,"rss_mib":1154.0078125},{"t_s":48,"rss_mib":1154.02734375},{"t_s":48,"rss_mib":1154.02734375},{"t_s":48,"rss_mib":1154.03515625},{"t_s":48,"rss_mib":1154.03515625},{"t_s":49,"rss_mib":1154.1484375},{"t_s":49,"rss_mib":1154.15625},{"t_s":50,"rss_mib":1154.15625},{"t_s":50,"rss_mib":1154.15625},{"t_s":50,"rss_mib":1154.15625},{"t_s":51,"rss_mib":1154.15625},{"t_s":51,"rss_mib":1154.15625},{"t_s":52,"rss_mib":1154.17578125},{"t_s":52,"rss_mib":1154.17578125},{"t_s":53,"rss_mib":1154.17578125},{"t_s":53,"rss_mib":1154.17578125},{"t_s":53,"rss_mib":1154.17578125},{"t_s":54,"rss_mib":1154.1875},{"t_s":54,"rss_mib":1154.1875},{"t_s":54,"rss_mib":1154.1875},{"t_s":55,"rss_mib":1154.20703125},{"t_s":55,"rss_mib":1154.20703125},{"t_s":55,"rss_mib":1154.21484375},{"t_s":55,"rss_mib":1154.21484375},{"t_s":56,"rss_mib":1154.21484375},{"t_s":56,"rss_mib":1154.21484375},{"t_s":57,"rss_mib":1154.21875},{"t_s":57,"rss_mib":1154.21875},{"t_s":57,"rss_mib":1154.23046875},{"t_s":57,"rss_mib":1154.23046875},{"t_s":58,"rss_mib":1154.23046875},{"t_s":58,"rss_mib":1154.23046875},{"t_s":59,"rss_mib":1154.23828125},{"t_s":59,"rss_mib":1154.23828125},{"t_s":59,"rss_mib":1154.23828125}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.009898073,"cost":6.034862998,"memory":438.912216565,"streaming":2.679785033,"streams_sustained":528.771034495,"throughput":321.802818215},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"}}}}},"openai-responses":{"configurable":true,"served":true,"egress_config":null,"serve_error":"","cells":{"openai":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1/chat/completions","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"gemini":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1beta/models/gpt-4o-mini:generateContent","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"cohere":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v2/chat","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"bedrock":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/model/gpt-4o-mini/converse","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"openai-responses":{"served":true,"reason":null,"status":"","path":"/v1/responses","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":642,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":747,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":775,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":633.7934164508637,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1577.8011794439765,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.9859729286272246,"certified":true,"suppressed":false},"cost_window_ok":{"value":37457,"certified":true,"suppressed":false},"cost_window_rps":{"value":6242,"certified":true,"suppressed":false},"cost_threads":{"value":141,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.0010144966227941374,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":2925,"certified":true,"suppressed":false},"concurrency":2,"p99_us":890,"first_disqualified_conc":4,"lower_bound":false,"samples":17553},{"bound_ms":5,"rps":{"value":6438,"certified":true,"suppressed":false},"concurrency":16,"p99_us":3521,"first_disqualified_conc":32,"lower_bound":false,"samples":38636},{"bound_ms":10,"rps":{"value":6717,"certified":true,"suppressed":false},"concurrency":32,"p99_us":6634,"first_disqualified_conc":64,"lower_bound":false,"samples":40311},{"bound_ms":50,"rps":{"value":6791,"certified":true,"suppressed":false},"concurrency":128,"p99_us":26658,"first_disqualified_conc":256,"lower_bound":false,"samples":40754},{"bound_ms":100,"rps":{"value":6791,"certified":true,"suppressed":false},"concurrency":128,"p99_us":26658,"first_disqualified_conc":512,"lower_bound":false,"samples":40754},{"bound_ms":null,"rps":{"value":6791,"certified":true,"suppressed":false},"concurrency":128,"p99_us":26658,"first_disqualified_conc":null,"lower_bound":false,"samples":40754}],"sweep":[{"conc":1,"ok":8763,"rps":1460,"p99_us":895,"fail":0},{"conc":1,"ok":8782,"rps":1463,"p99_us":894,"fail":0},{"conc":1,"ok":8945,"rps":1490,"p99_us":889,"fail":0},{"conc":2,"ok":9813,"rps":1635,"p99_us":1512,"fail":0},{"conc":2,"ok":17553,"rps":2925,"p99_us":890,"fail":0},{"conc":2,"ok":17256,"rps":2875,"p99_us":902,"fail":0},{"conc":4,"ok":19125,"rps":3186,"p99_us":1563,"fail":0},{"conc":4,"ok":27290,"rps":4547,"p99_us":1468,"fail":0},{"conc":4,"ok":27248,"rps":4540,"p99_us":1471,"fail":0},{"conc":8,"ok":28302,"rps":4716,"p99_us":3611,"fail":0},{"conc":8,"ok":36496,"rps":6081,"p99_us":2807,"fail":0},{"conc":8,"ok":28301,"rps":4716,"p99_us":3440,"fail":0},{"conc":16,"ok":38636,"rps":6438,"p99_us":3521,"fail":0},{"conc":16,"ok":38337,"rps":6388,"p99_us":3619,"fail":0},{"conc":16,"ok":38327,"rps":6386,"p99_us":3999,"fail":0},{"conc":32,"ok":40161,"rps":6692,"p99_us":7313,"fail":0},{"conc":32,"ok":40311,"rps":6717,"p99_us":6634,"fail":0},{"conc":32,"ok":39908,"rps":6651,"p99_us":8279,"fail":0},{"conc":64,"ok":40530,"rps":6753,"p99_us":13131,"fail":0},{"conc":64,"ok":40705,"rps":6783,"p99_us":13229,"fail":0},{"conc":64,"ok":40705,"rps":6783,"p99_us":14349,"fail":0},{"conc":128,"ok":40718,"rps":6785,"p99_us":26588,"fail":0},{"conc":128,"ok":40709,"rps":6783,"p99_us":25185,"fail":0},{"conc":128,"ok":40754,"rps":6791,"p99_us":26658,"fail":0},{"conc":256,"ok":39693,"rps":6613,"p99_us":50179,"fail":0},{"conc":256,"ok":39505,"rps":6583,"p99_us":50796,"fail":0},{"conc":256,"ok":39542,"rps":6589,"p99_us":53597,"fail":0},{"conc":512,"ok":37755,"rps":6292,"p99_us":107608,"fail":0},{"conc":512,"ok":37618,"rps":6268,"p99_us":100940,"fail":0},{"conc":512,"ok":37709,"rps":6283,"p99_us":102809,"fail":0},{"conc":1024,"ok":37374,"rps":6227,"p99_us":217563,"fail":0},{"conc":1024,"ok":37394,"rps":6230,"p99_us":203520,"fail":0},{"conc":1024,"ok":37347,"rps":6223,"p99_us":208089,"fail":0},{"conc":2048,"ok":37879,"rps":6311,"p99_us":920445,"fail":0},{"conc":2048,"ok":37961,"rps":6325,"p99_us":814459,"fail":0},{"conc":2048,"ok":38054,"rps":6341,"p99_us":632104,"fail":0},{"conc":4096,"ok":38110,"rps":6351,"p99_us":552889,"fail":0},{"conc":4096,"ok":38091,"rps":6347,"p99_us":759535,"fail":0},{"conc":4096,"ok":38047,"rps":6340,"p99_us":673225,"fail":0},{"conc":8192,"ok":38009,"rps":6333,"p99_us":819647,"fail":0},{"conc":8192,"ok":37473,"rps":6244,"p99_us":740999,"fail":0},{"conc":8192,"ok":37235,"rps":6204,"p99_us":875860,"fail":0},{"conc":16384,"ok":37511,"rps":6251,"p99_us":740744,"fail":0},{"conc":16384,"ok":37675,"rps":6277,"p99_us":489583,"fail":0},{"conc":16384,"ok":37948,"rps":6323,"p99_us":520661,"fail":0},{"conc":32768,"ok":29694,"rps":4948,"p99_us":3158971,"fail":0},{"conc":32768,"ok":29627,"rps":4936,"p99_us":2724411,"fail":0},{"conc":32768,"ok":31454,"rps":5241,"p99_us":2690523,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1463,"certified":true,"suppressed":false},"p99_us":894,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":2875,"certified":true,"suppressed":false},"p99_us":902,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":4540,"certified":true,"suppressed":false},"p99_us":1471,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":4716,"certified":true,"suppressed":false},"p99_us":3440,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":6388,"certified":true,"suppressed":false},"p99_us":3619,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":6692,"certified":true,"suppressed":false},"p99_us":7313,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":6783,"certified":true,"suppressed":false},"p99_us":13229,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":6785,"certified":true,"suppressed":false},"p99_us":26588,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":6589,"certified":true,"suppressed":false},"p99_us":50796,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":6283,"certified":true,"suppressed":false},"p99_us":102809,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":6227,"certified":true,"suppressed":false},"p99_us":208089,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":6325,"certified":true,"suppressed":false},"p99_us":814459,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":6347,"certified":true,"suppressed":false},"p99_us":673225,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":6244,"certified":true,"suppressed":false},"p99_us":819647,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":6277,"certified":true,"suppressed":false},"p99_us":520661,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":4948,"certified":true,"suppressed":false},"p99_us":2724411,"windows":3,"clean_windows":3}],"reverify_note":"openai-responses>openai-responses is a same-dialect cell: there is no translation to prove, so a request-shape check upstream cannot distinguish a translating gateway from a forwarding one","c1_note":"the c=1 percentiles are taken over 8971 successful gateway round trip(s) and 260207 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":8971,"direct_c1_samples":260207},"stream":{"stream_served":"untestable","stream_error":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here","reason":"untestable","stream_c1_note":null,"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"}},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1147.16796875,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1156.55078125,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1157.92578125,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1165.46875,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1186.5546875,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":66,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.5987504753473016,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":0.6971452873324683,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1147.84375},{"t_s":2,"rss_mib":1164.21484375},{"t_s":3,"rss_mib":1164.21484375},{"t_s":5,"rss_mib":1164.28515625},{"t_s":5,"rss_mib":1164.28515625},{"t_s":8,"rss_mib":1164.5859375},{"t_s":8,"rss_mib":1164.5859375},{"t_s":11,"rss_mib":1164.6328125},{"t_s":11,"rss_mib":1164.6328125},{"t_s":14,"rss_mib":1164.80078125},{"t_s":14,"rss_mib":1164.80078125},{"t_s":17,"rss_mib":1164.84375},{"t_s":17,"rss_mib":1164.84375},{"t_s":20,"rss_mib":1164.8828125},{"t_s":20,"rss_mib":1164.8828125},{"t_s":23,"rss_mib":1164.90625},{"t_s":23,"rss_mib":1164.90625},{"t_s":26,"rss_mib":1164.953125},{"t_s":26,"rss_mib":1164.953125},{"t_s":28,"rss_mib":1165.015625},{"t_s":29,"rss_mib":1165.0234375},{"t_s":32,"rss_mib":1165.06640625},{"t_s":32,"rss_mib":1165.06640625},{"t_s":33,"rss_mib":1165.07421875},{"t_s":35,"rss_mib":1165.08203125},{"t_s":37,"rss_mib":1165.0859375},{"t_s":38,"rss_mib":1165.0859375},{"t_s":41,"rss_mib":1165.10546875},{"t_s":41,"rss_mib":1165.10546875},{"t_s":43,"rss_mib":1165.12109375},{"t_s":44,"rss_mib":1165.12109375},{"t_s":46,"rss_mib":1165.12890625},{"t_s":47,"rss_mib":1165.12890625},{"t_s":50,"rss_mib":1165.35546875},{"t_s":50,"rss_mib":1165.35546875},{"t_s":53,"rss_mib":1165.43359375},{"t_s":54,"rss_mib":1165.46875},{"t_s":55,"rss_mib":1161.4609375},{"t_s":59,"rss_mib":1161.62109375},{"t_s":59,"rss_mib":1160.52734375},{"t_s":60,"rss_mib":1160.52734375},{"t_s":62,"rss_mib":1160.62890625},{"t_s":63,"rss_mib":1160.62890625},{"t_s":64,"rss_mib":1160.63671875},{"t_s":66,"rss_mib":1160.640625},{"t_s":69,"rss_mib":1160.640625},{"t_s":72,"rss_mib":1160.640625},{"t_s":75,"rss_mib":1160.640625},{"t_s":78,"rss_mib":1160.640625},{"t_s":81,"rss_mib":1160.640625},{"t_s":83,"rss_mib":1160.64453125},{"t_s":84,"rss_mib":1160.64453125},{"t_s":86,"rss_mib":1160.6640625},{"t_s":87,"rss_mib":1160.6640625},{"t_s":89,"rss_mib":1160.85546875},{"t_s":90,"rss_mib":1160.85546875},{"t_s":93,"rss_mib":1160.85546875},{"t_s":95,"rss_mib":1160.86328125},{"t_s":96,"rss_mib":1160.86328125},{"t_s":99,"rss_mib":1160.86328125},{"t_s":101,"rss_mib":1160.87890625},{"t_s":102,"rss_mib":1160.87890625},{"t_s":105,"rss_mib":1160.87890625},{"t_s":108,"rss_mib":1160.87890625},{"t_s":111,"rss_mib":1160.87890625},{"t_s":113,"rss_mib":1160.96875},{"t_s":114,"rss_mib":1160.96875},{"t_s":117,"rss_mib":1160.96875},{"t_s":119,"rss_mib":1160.98046875},{"t_s":120,"rss_mib":1160.984375},{"t_s":120,"rss_mib":1160.98828125},{"t_s":123,"rss_mib":1160.99609375},{"t_s":124,"rss_mib":1161.00390625},{"t_s":126,"rss_mib":1161.00390625},{"t_s":128,"rss_mib":1161.015625},{"t_s":129,"rss_mib":1161.015625},{"t_s":131,"rss_mib":1161.02734375},{"t_s":132,"rss_mib":1161.04296875},{"t_s":134,"rss_mib":1161.14453125},{"t_s":135,"rss_mib":1161.1796875},{"t_s":135,"rss_mib":1161.1875},{"t_s":138,"rss_mib":1161.1875},{"t_s":139,"rss_mib":1161.1953125},{"t_s":141,"rss_mib":1161.1953125},{"t_s":142,"rss_mib":1161.203125},{"t_s":144,"rss_mib":1161.203125},{"t_s":147,"rss_mib":1161.203125},{"t_s":147,"rss_mib":1161.25},{"t_s":150,"rss_mib":1161.25},{"t_s":153,"rss_mib":1161.25},{"t_s":154,"rss_mib":1161.2578125},{"t_s":156,"rss_mib":1161.2578125},{"t_s":159,"rss_mib":1161.2578125},{"t_s":161,"rss_mib":1161.2734375},{"t_s":162,"rss_mib":1161.2734375},{"t_s":165,"rss_mib":1161.2734375},{"t_s":168,"rss_mib":1161.29296875},{"t_s":168,"rss_mib":1161.3046875},{"t_s":171,"rss_mib":1161.3046875},{"t_s":173,"rss_mib":1161.359375},{"t_s":174,"rss_mib":1161.359375},{"t_s":174,"rss_mib":1158.93359375},{"t_s":179,"rss_mib":1158.96875},{"t_s":179,"rss_mib":1157.73046875},{"t_s":180,"rss_mib":1157.73828125},{"t_s":181,"rss_mib":1157.8671875},{"t_s":183,"rss_mib":1157.8671875},{"t_s":186,"rss_mib":1157.87890625},{"t_s":186,"rss_mib":1157.87890625},{"t_s":189,"rss_mib":1157.87890625},{"t_s":189,"rss_mib":1156.1015625},{"t_s":192,"rss_mib":1156.171875},{"t_s":195,"rss_mib":1156.171875},{"t_s":198,"rss_mib":1156.171875},{"t_s":201,"rss_mib":1156.171875},{"t_s":201,"rss_mib":1156.1796875},{"t_s":204,"rss_mib":1156.1796875},{"t_s":207,"rss_mib":1156.20703125},{"t_s":207,"rss_mib":1156.20703125},{"t_s":209,"rss_mib":1156.296875},{"t_s":210,"rss_mib":1156.296875},{"t_s":213,"rss_mib":1156.296875},{"t_s":215,"rss_mib":1156.30859375},{"t_s":216,"rss_mib":1156.32421875},{"t_s":219,"rss_mib":1156.32421875},{"t_s":221,"rss_mib":1156.3359375},{"t_s":222,"rss_mib":1156.3359375},{"t_s":225,"rss_mib":1156.3359375},{"t_s":228,"rss_mib":1156.3359375},{"t_s":231,"rss_mib":1156.3359375},{"t_s":234,"rss_mib":1156.4296875},{"t_s":234,"rss_mib":1156.4296875},{"t_s":235,"rss_mib":1156.4375},{"t_s":237,"rss_mib":1156.4375},{"t_s":240,"rss_mib":1156.4375},{"t_s":242,"rss_mib":1156.4453125},{"t_s":243,"rss_mib":1156.4453125},{"t_s":246,"rss_mib":1156.4453125},{"t_s":249,"rss_mib":1156.4453125},{"t_s":252,"rss_mib":1156.4453125},{"t_s":254,"rss_mib":1156.515625},{"t_s":255,"rss_mib":1156.515625},{"t_s":257,"rss_mib":1156.5234375},{"t_s":258,"rss_mib":1156.5234375},{"t_s":259,"rss_mib":1156.53515625},{"t_s":261,"rss_mib":1156.53515625},{"t_s":263,"rss_mib":1156.55078125},{"t_s":264,"rss_mib":1156.55078125},{"t_s":267,"rss_mib":1156.55078125},{"t_s":267,"rss_mib":1156.640625},{"t_s":270,"rss_mib":1156.640625},{"t_s":273,"rss_mib":1156.640625},{"t_s":273,"rss_mib":1156.79296875},{"t_s":276,"rss_mib":1156.79296875},{"t_s":277,"rss_mib":1156.8046875},{"t_s":279,"rss_mib":1156.8046875},{"t_s":282,"rss_mib":1156.8046875},{"t_s":282,"rss_mib":1156.80859375},{"t_s":285,"rss_mib":1156.80859375},{"t_s":288,"rss_mib":1156.80859375},{"t_s":290,"rss_mib":1156.8125},{"t_s":291,"rss_mib":1156.8125},{"t_s":293,"rss_mib":1156.921875},{"t_s":294,"rss_mib":1156.921875},{"t_s":294,"rss_mib":1156.9609375},{"t_s":297,"rss_mib":1156.9609375},{"t_s":299,"rss_mib":1156.96875},{"t_s":300,"rss_mib":1156.96875},{"t_s":301,"rss_mib":1156.97265625},{"t_s":303,"rss_mib":1156.97265625},{"t_s":306,"rss_mib":1156.9765625},{"t_s":306,"rss_mib":1156.9765625},{"t_s":307,"rss_mib":1156.984375},{"t_s":309,"rss_mib":1156.984375},{"t_s":312,"rss_mib":1156.984375},{"t_s":315,"rss_mib":1156.984375},{"t_s":317,"rss_mib":1157.0078125},{"t_s":318,"rss_mib":1157.0078125},{"t_s":321,"rss_mib":1157.0078125},{"t_s":324,"rss_mib":1157.0078125},{"t_s":325,"rss_mib":1157.015625},{"t_s":327,"rss_mib":1157.015625},{"t_s":327,"rss_mib":1157.80078125},{"t_s":330,"rss_mib":1157.80078125},{"t_s":333,"rss_mib":1157.80078125},{"t_s":334,"rss_mib":1157.8046875},{"t_s":336,"rss_mib":1157.8046875},{"t_s":339,"rss_mib":1157.83203125},{"t_s":339,"rss_mib":1157.91796875},{"t_s":340,"rss_mib":1157.921875},{"t_s":342,"rss_mib":1157.921875},{"t_s":344,"rss_mib":1157.92578125},{"t_s":345,"rss_mib":1157.92578125},{"t_s":348,"rss_mib":1157.93359375},{"t_s":348,"rss_mib":1157.93359375},{"t_s":350,"rss_mib":1157.94140625},{"t_s":351,"rss_mib":1157.94140625},{"t_s":353,"rss_mib":1158.02734375},{"t_s":354,"rss_mib":1158.02734375},{"t_s":357,"rss_mib":1158.02734375},{"t_s":360,"rss_mib":1158.02734375}],"idle_rss_series":[{"t_s":0,"rss_mib":1146.68359375},{"t_s":0,"rss_mib":1146.74609375},{"t_s":0,"rss_mib":1146.74609375},{"t_s":0,"rss_mib":1146.74609375},{"t_s":1,"rss_mib":1146.74609375},{"t_s":1,"rss_mib":1146.8515625},{"t_s":1,"rss_mib":1146.8515625},{"t_s":2,"rss_mib":1146.85546875},{"t_s":2,"rss_mib":1146.85546875},{"t_s":2,"rss_mib":1146.85546875},{"t_s":3,"rss_mib":1146.85546875},{"t_s":3,"rss_mib":1146.85546875},{"t_s":4,"rss_mib":1146.85546875},{"t_s":4,"rss_mib":1146.85546875},{"t_s":5,"rss_mib":1146.85546875},{"t_s":5,"rss_mib":1146.85546875},{"t_s":6,"rss_mib":1146.85546875},{"t_s":6,"rss_mib":1146.85546875},{"t_s":7,"rss_mib":1146.859375},{"t_s":7,"rss_mib":1146.859375},{"t_s":7,"rss_mib":1146.859375},{"t_s":8,"rss_mib":1146.859375},{"t_s":8,"rss_mib":1146.859375},{"t_s":9,"rss_mib":1146.859375},{"t_s":10,"rss_mib":1146.859375},{"t_s":10,"rss_mib":1146.859375},{"t_s":10,"rss_mib":1146.859375},{"t_s":11,"rss_mib":1146.86328125},{"t_s":11,"rss_mib":1146.86328125},{"t_s":11,"rss_mib":1146.86328125},{"t_s":12,"rss_mib":1146.86328125},{"t_s":13,"rss_mib":1146.86328125},{"t_s":13,"rss_mib":1146.8671875},{"t_s":13,"rss_mib":1146.8671875},{"t_s":13,"rss_mib":1146.8671875},{"t_s":14,"rss_mib":1146.8671875},{"t_s":14,"rss_mib":1146.8671875},{"t_s":15,"rss_mib":1146.8671875},{"t_s":15,"rss_mib":1146.8671875},{"t_s":16,"rss_mib":1146.8671875},{"t_s":16,"rss_mib":1146.8671875},{"t_s":17,"rss_mib":1146.8671875},{"t_s":17,"rss_mib":1146.8671875},{"t_s":18,"rss_mib":1146.875},{"t_s":18,"rss_mib":1146.875},{"t_s":18,"rss_mib":1146.875},{"t_s":19,"rss_mib":1146.875},{"t_s":20,"rss_mib":1146.875},{"t_s":20,"rss_mib":1146.875},{"t_s":20,"rss_mib":1146.875},{"t_s":21,"rss_mib":1146.875},{"t_s":21,"rss_mib":1146.90234375},{"t_s":21,"rss_mib":1146.90234375},{"t_s":22,"rss_mib":1146.91015625},{"t_s":22,"rss_mib":1146.91015625},{"t_s":23,"rss_mib":1146.91015625},{"t_s":23,"rss_mib":1146.91015625},{"t_s":23,"rss_mib":1146.91796875},{"t_s":24,"rss_mib":1146.91796875},{"t_s":24,"rss_mib":1146.9765625},{"t_s":24,"rss_mib":1146.9765625},{"t_s":25,"rss_mib":1146.9921875},{"t_s":25,"rss_mib":1146.9921875},{"t_s":26,"rss_mib":1146.9921875},{"t_s":26,"rss_mib":1147},{"t_s":26,"rss_mib":1147},{"t_s":26,"rss_mib":1147},{"t_s":27,"rss_mib":1147},{"t_s":27,"rss_mib":1147.15625},{"t_s":27,"rss_mib":1147.15625},{"t_s":28,"rss_mib":1147.16015625},{"t_s":28,"rss_mib":1147.16015625},{"t_s":28,"rss_mib":1147.16796875},{"t_s":28,"rss_mib":1147.16796875},{"t_s":29,"rss_mib":1147.16796875},{"t_s":30,"rss_mib":1147.16796875},{"t_s":30,"rss_mib":1147.1796875},{"t_s":30,"rss_mib":1147.1796875},{"t_s":30,"rss_mib":1147.1875},{"t_s":31,"rss_mib":1147.1875},{"t_s":31,"rss_mib":1147.1875},{"t_s":32,"rss_mib":1147.1875},{"t_s":33,"rss_mib":1147.1875},{"t_s":33,"rss_mib":1147.1875},{"t_s":33,"rss_mib":1147.1953125},{"t_s":33,"rss_mib":1147.1953125},{"t_s":34,"rss_mib":1147.1953125},{"t_s":34,"rss_mib":1147.1953125},{"t_s":35,"rss_mib":1147.203125},{"t_s":35,"rss_mib":1147.203125},{"t_s":36,"rss_mib":1147.203125},{"t_s":36,"rss_mib":1147.203125},{"t_s":36,"rss_mib":1147.203125},{"t_s":37,"rss_mib":1147.2109375},{"t_s":37,"rss_mib":1147.2109375},{"t_s":37,"rss_mib":1147.2109375},{"t_s":38,"rss_mib":1147.2109375},{"t_s":38,"rss_mib":1147.21875},{"t_s":39,"rss_mib":1147.21875},{"t_s":39,"rss_mib":1147.21875},{"t_s":40,"rss_mib":1147.21875},{"t_s":40,"rss_mib":1147.21875},{"t_s":40,"rss_mib":1147.21875},{"t_s":41,"rss_mib":1147.2265625},{"t_s":41,"rss_mib":1147.2265625},{"t_s":41,"rss_mib":1147.2265625},{"t_s":42,"rss_mib":1147.2265625},{"t_s":43,"rss_mib":1147.23828125},{"t_s":43,"rss_mib":1147.23828125},{"t_s":43,"rss_mib":1147.23828125},{"t_s":44,"rss_mib":1147.23828125},{"t_s":44,"rss_mib":1147.24609375},{"t_s":44,"rss_mib":1147.24609375},{"t_s":45,"rss_mib":1147.24609375},{"t_s":46,"rss_mib":1147.24609375},{"t_s":46,"rss_mib":1147.25},{"t_s":46,"rss_mib":1147.25},{"t_s":46,"rss_mib":1147.2578125},{"t_s":47,"rss_mib":1147.265625},{"t_s":47,"rss_mib":1147.265625},{"t_s":47,"rss_mib":1147.265625},{"t_s":48,"rss_mib":1147.26953125},{"t_s":48,"rss_mib":1147.26953125},{"t_s":49,"rss_mib":1147.26953125},{"t_s":49,"rss_mib":1147.27734375},{"t_s":49,"rss_mib":1147.27734375},{"t_s":50,"rss_mib":1147.27734375},{"t_s":50,"rss_mib":1147.27734375},{"t_s":51,"rss_mib":1147.28515625},{"t_s":51,"rss_mib":1147.28515625},{"t_s":51,"rss_mib":1147.29296875},{"t_s":52,"rss_mib":1147.29296875},{"t_s":52,"rss_mib":1147.3046875},{"t_s":52,"rss_mib":1147.3046875},{"t_s":52,"rss_mib":1147.3125},{"t_s":53,"rss_mib":1147.3125},{"t_s":53,"rss_mib":1147.328125},{"t_s":53,"rss_mib":1147.41015625},{"t_s":53,"rss_mib":1147.41015625},{"t_s":54,"rss_mib":1147.41015625},{"t_s":54,"rss_mib":1147.41015625},{"t_s":55,"rss_mib":1147.4140625},{"t_s":55,"rss_mib":1147.4140625},{"t_s":56,"rss_mib":1147.4140625},{"t_s":56,"rss_mib":1147.41796875},{"t_s":56,"rss_mib":1147.41796875},{"t_s":56,"rss_mib":1147.41796875},{"t_s":57,"rss_mib":1147.42578125},{"t_s":57,"rss_mib":1147.42578125},{"t_s":57,"rss_mib":1147.42578125},{"t_s":58,"rss_mib":1147.42578125},{"t_s":58,"rss_mib":1147.43359375},{"t_s":59,"rss_mib":1147.43359375},{"t_s":59,"rss_mib":1147.4375},{"t_s":59,"rss_mib":1147.4375},{"t_s":59,"rss_mib":1147.4375}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.009846491,"cost":6.034163247,"memory":434.966872496,"streaming":0.000001947,"streams_sustained":0.000001657,"throughput":319.826568735},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"stream.added_gap_p50_us":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_gap_p99_us":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p50_us":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p99_us":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_direct_samples":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_gw_samples":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_fps":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_headroom":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_mock_ceiling":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_direct_samples":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_gw_samples":{"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here"}}},"anthropic":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1/messages","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}}}},"gemini":{"configurable":true,"served":true,"egress_config":null,"serve_error":"","cells":{"openai":{"served":true,"reason":null,"status":"","path":"/v1/chat/completions","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":923,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":1226,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":1254,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":951.6544794570992,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1050.8015478164732,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.7511344229974055,"certified":true,"suppressed":false},"cost_window_ok":{"value":19009,"certified":true,"suppressed":false},"cost_window_rps":{"value":3167,"certified":true,"suppressed":false},"cost_threads":{"value":141,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.011047398600662844,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":4171,"certified":true,"suppressed":false},"concurrency":8,"p99_us":4447,"first_disqualified_conc":16,"lower_bound":false,"samples":25035},{"bound_ms":10,"rps":{"value":4327,"certified":true,"suppressed":false},"concurrency":16,"p99_us":6571,"first_disqualified_conc":32,"lower_bound":false,"samples":25969},{"bound_ms":50,"rps":{"value":4449,"certified":true,"suppressed":false},"concurrency":64,"p99_us":22764,"first_disqualified_conc":256,"lower_bound":false,"samples":26699},{"bound_ms":100,"rps":{"value":4449,"certified":true,"suppressed":false},"concurrency":64,"p99_us":22764,"first_disqualified_conc":512,"lower_bound":false,"samples":26699},{"bound_ms":null,"rps":{"value":4449,"certified":true,"suppressed":false},"concurrency":64,"p99_us":22764,"first_disqualified_conc":null,"lower_bound":false,"samples":26699}],"sweep":[{"conc":1,"ok":6101,"rps":1016,"p99_us":1257,"fail":0},{"conc":1,"ok":6090,"rps":1014,"p99_us":1268,"fail":0},{"conc":1,"ok":6230,"rps":1038,"p99_us":1248,"fail":0},{"conc":2,"ok":11965,"rps":1993,"p99_us":1283,"fail":0},{"conc":2,"ok":12076,"rps":2011,"p99_us":1285,"fail":0},{"conc":2,"ok":6633,"rps":1105,"p99_us":2254,"fail":0},{"conc":4,"ok":18723,"rps":3120,"p99_us":2268,"fail":0},{"conc":4,"ok":18676,"rps":3111,"p99_us":2268,"fail":0},{"conc":4,"ok":18611,"rps":3101,"p99_us":2311,"fail":0},{"conc":8,"ok":24989,"rps":4164,"p99_us":3484,"fail":0},{"conc":8,"ok":25035,"rps":4171,"p99_us":4447,"fail":0},{"conc":8,"ok":24683,"rps":4113,"p99_us":5231,"fail":0},{"conc":16,"ok":25737,"rps":4288,"p99_us":7757,"fail":0},{"conc":16,"ok":25769,"rps":4294,"p99_us":6739,"fail":0},{"conc":16,"ok":25969,"rps":4327,"p99_us":6571,"fail":0},{"conc":32,"ok":26356,"rps":4392,"p99_us":14907,"fail":0},{"conc":32,"ok":26587,"rps":4430,"p99_us":12045,"fail":0},{"conc":32,"ok":26491,"rps":4414,"p99_us":12128,"fail":0},{"conc":64,"ok":26541,"rps":4422,"p99_us":21059,"fail":0},{"conc":64,"ok":26673,"rps":4444,"p99_us":21278,"fail":0},{"conc":64,"ok":26699,"rps":4449,"p99_us":22764,"fail":0},{"conc":128,"ok":26594,"rps":4431,"p99_us":37904,"fail":0},{"conc":128,"ok":26543,"rps":4423,"p99_us":37191,"fail":0},{"conc":128,"ok":26431,"rps":4404,"p99_us":41000,"fail":0},{"conc":256,"ok":25549,"rps":4257,"p99_us":78161,"fail":0},{"conc":256,"ok":25486,"rps":4247,"p99_us":78021,"fail":0},{"conc":256,"ok":25520,"rps":4252,"p99_us":76379,"fail":0},{"conc":512,"ok":25000,"rps":4166,"p99_us":161260,"fail":0},{"conc":512,"ok":24947,"rps":4157,"p99_us":155614,"fail":0},{"conc":512,"ok":25087,"rps":4179,"p99_us":149850,"fail":0},{"conc":1024,"ok":25407,"rps":4234,"p99_us":452518,"fail":0},{"conc":1024,"ok":25405,"rps":4232,"p99_us":363430,"fail":0},{"conc":1024,"ok":25405,"rps":4233,"p99_us":2104156,"fail":0},{"conc":2048,"ok":25060,"rps":4175,"p99_us":1073657,"fail":0},{"conc":2048,"ok":24887,"rps":4147,"p99_us":1022029,"fail":0},{"conc":2048,"ok":25041,"rps":4172,"p99_us":1136469,"fail":0},{"conc":4096,"ok":25141,"rps":4189,"p99_us":1619062,"fail":0},{"conc":4096,"ok":24967,"rps":4160,"p99_us":924064,"fail":0},{"conc":4096,"ok":24985,"rps":4163,"p99_us":1423440,"fail":0},{"conc":8192,"ok":24928,"rps":4154,"p99_us":1210903,"fail":0},{"conc":8192,"ok":24984,"rps":4163,"p99_us":982916,"fail":0},{"conc":8192,"ok":24907,"rps":4150,"p99_us":865542,"fail":0},{"conc":16384,"ok":24907,"rps":4150,"p99_us":956774,"fail":0},{"conc":16384,"ok":25193,"rps":4198,"p99_us":1064589,"fail":0},{"conc":16384,"ok":25202,"rps":4199,"p99_us":1112987,"fail":0},{"conc":32768,"ok":22499,"rps":3748,"p99_us":2821530,"fail":0},{"conc":32768,"ok":22011,"rps":3666,"p99_us":2591714,"fail":0},{"conc":32768,"ok":23316,"rps":3885,"p99_us":2629202,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1016,"certified":true,"suppressed":false},"p99_us":1257,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":1993,"certified":true,"suppressed":false},"p99_us":1285,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":3111,"certified":true,"suppressed":false},"p99_us":2268,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":4164,"certified":true,"suppressed":false},"p99_us":4447,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":4294,"certified":true,"suppressed":false},"p99_us":6739,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":4414,"certified":true,"suppressed":false},"p99_us":12128,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":4444,"certified":true,"suppressed":false},"p99_us":21278,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":4423,"certified":true,"suppressed":false},"p99_us":37904,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":4252,"certified":true,"suppressed":false},"p99_us":78021,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":4166,"certified":true,"suppressed":false},"p99_us":155614,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":4233,"certified":true,"suppressed":false},"p99_us":452518,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":4172,"certified":true,"suppressed":false},"p99_us":1073657,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":4163,"certified":true,"suppressed":false},"p99_us":1423440,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":4154,"certified":true,"suppressed":false},"p99_us":982916,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":4198,"certified":true,"suppressed":false},"p99_us":1064589,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":3748,"certified":true,"suppressed":false},"p99_us":2629202,"windows":3,"clean_windows":3}],"egress_reverified":true,"c1_note":"the c=1 percentiles are taken over 6239 successful gateway round trip(s) and 265561 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":6239,"direct_c1_samples":265561},"stream":{"stream_served":"untestable","stream_error":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","reason":"untestable","stream_c1_note":null,"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"}},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1150.5859375,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1160.25390625,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1160.609375,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1170.90234375,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1268.890625,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.25904735694337694,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":0.6054154815784856,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1151.25390625},{"t_s":1,"rss_mib":1168.015625},{"t_s":2,"rss_mib":1168.015625},{"t_s":5,"rss_mib":1168.0859375},{"t_s":5,"rss_mib":1168.08984375},{"t_s":8,"rss_mib":1168.1796875},{"t_s":8,"rss_mib":1168.1796875},{"t_s":11,"rss_mib":1168.23046875},{"t_s":11,"rss_mib":1168.23046875},{"t_s":13,"rss_mib":1168.47265625},{"t_s":14,"rss_mib":1168.47265625},{"t_s":17,"rss_mib":1168.52734375},{"t_s":17,"rss_mib":1168.52734375},{"t_s":19,"rss_mib":1168.65234375},{"t_s":20,"rss_mib":1168.65234375},{"t_s":22,"rss_mib":1168.68359375},{"t_s":23,"rss_mib":1168.68359375},{"t_s":25,"rss_mib":1168.78515625},{"t_s":26,"rss_mib":1168.78515625},{"t_s":29,"rss_mib":1168.80078125},{"t_s":29,"rss_mib":1168.80078125},{"t_s":32,"rss_mib":1168.8359375},{"t_s":32,"rss_mib":1168.8359375},{"t_s":35,"rss_mib":1168.859375},{"t_s":35,"rss_mib":1168.859375},{"t_s":38,"rss_mib":1168.88671875},{"t_s":38,"rss_mib":1168.88671875},{"t_s":41,"rss_mib":1168.984375},{"t_s":42,"rss_mib":1168.984375},{"t_s":44,"rss_mib":1169.1171875},{"t_s":45,"rss_mib":1169.1171875},{"t_s":47,"rss_mib":1169.2734375},{"t_s":48,"rss_mib":1169.2734375},{"t_s":50,"rss_mib":1169.33984375},{"t_s":51,"rss_mib":1169.33984375},{"t_s":54,"rss_mib":1169.33984375},{"t_s":55,"rss_mib":1169.79296875},{"t_s":57,"rss_mib":1169.79296875},{"t_s":59,"rss_mib":1169.91015625},{"t_s":60,"rss_mib":1169.91015625},{"t_s":60,"rss_mib":1169.93359375},{"t_s":63,"rss_mib":1169.93359375},{"t_s":63,"rss_mib":1169.9375},{"t_s":66,"rss_mib":1169.9375},{"t_s":67,"rss_mib":1169.94140625},{"t_s":69,"rss_mib":1169.94140625},{"t_s":72,"rss_mib":1169.94140625},{"t_s":74,"rss_mib":1169.9453125},{"t_s":75,"rss_mib":1169.9453125},{"t_s":76,"rss_mib":1169.96484375},{"t_s":78,"rss_mib":1169.96484375},{"t_s":79,"rss_mib":1169.97265625},{"t_s":81,"rss_mib":1169.97265625},{"t_s":82,"rss_mib":1169.984375},{"t_s":84,"rss_mib":1169.984375},{"t_s":86,"rss_mib":1170.01171875},{"t_s":87,"rss_mib":1170.01171875},{"t_s":89,"rss_mib":1170.03125},{"t_s":90,"rss_mib":1170.03125},{"t_s":92,"rss_mib":1170.05078125},{"t_s":93,"rss_mib":1170.05078125},{"t_s":94,"rss_mib":1170.05859375},{"t_s":96,"rss_mib":1170.05859375},{"t_s":97,"rss_mib":1170.11328125},{"t_s":99,"rss_mib":1170.11328125},{"t_s":101,"rss_mib":1170.12890625},{"t_s":102,"rss_mib":1170.12890625},{"t_s":102,"rss_mib":1170.26171875},{"t_s":105,"rss_mib":1170.26171875},{"t_s":107,"rss_mib":1170.37109375},{"t_s":108,"rss_mib":1170.37109375},{"t_s":111,"rss_mib":1170.39453125},{"t_s":111,"rss_mib":1170.40625},{"t_s":113,"rss_mib":1170.41796875},{"t_s":114,"rss_mib":1170.421875},{"t_s":116,"rss_mib":1170.578125},{"t_s":117,"rss_mib":1170.58203125},{"t_s":119,"rss_mib":1170.59375},{"t_s":120,"rss_mib":1170.59375},{"t_s":122,"rss_mib":1170.62109375},{"t_s":123,"rss_mib":1170.62890625},{"t_s":126,"rss_mib":1170.65625},{"t_s":126,"rss_mib":1170.6640625},{"t_s":128,"rss_mib":1170.7109375},{"t_s":129,"rss_mib":1170.7109375},{"t_s":132,"rss_mib":1170.72265625},{"t_s":132,"rss_mib":1170.72265625},{"t_s":134,"rss_mib":1170.74609375},{"t_s":135,"rss_mib":1170.74609375},{"t_s":137,"rss_mib":1170.75390625},{"t_s":138,"rss_mib":1170.75390625},{"t_s":139,"rss_mib":1170.76171875},{"t_s":141,"rss_mib":1170.76171875},{"t_s":143,"rss_mib":1170.77734375},{"t_s":144,"rss_mib":1170.77734375},{"t_s":145,"rss_mib":1170.78515625},{"t_s":147,"rss_mib":1170.78515625},{"t_s":147,"rss_mib":1170.79296875},{"t_s":150,"rss_mib":1170.80078125},{"t_s":153,"rss_mib":1170.80078125},{"t_s":155,"rss_mib":1170.83203125},{"t_s":156,"rss_mib":1170.83203125},{"t_s":157,"rss_mib":1170.83984375},{"t_s":159,"rss_mib":1170.83984375},{"t_s":160,"rss_mib":1170.85546875},{"t_s":162,"rss_mib":1170.8828125},{"t_s":163,"rss_mib":1170.890625},{"t_s":165,"rss_mib":1170.90234375},{"t_s":167,"rss_mib":1165.28515625},{"t_s":168,"rss_mib":1165.28515625},{"t_s":168,"rss_mib":1165.22265625},{"t_s":171,"rss_mib":1165.22265625},{"t_s":174,"rss_mib":1165.37109375},{"t_s":174,"rss_mib":1161.6875},{"t_s":179,"rss_mib":1161.69921875},{"t_s":179,"rss_mib":1160.00390625},{"t_s":180,"rss_mib":1160.00390625},{"t_s":180,"rss_mib":1159.87890625},{"t_s":183,"rss_mib":1159.890625},{"t_s":184,"rss_mib":1159.9140625},{"t_s":186,"rss_mib":1159.9140625},{"t_s":189,"rss_mib":1159.9140625},{"t_s":192,"rss_mib":1159.9140625},{"t_s":194,"rss_mib":1159.92578125},{"t_s":195,"rss_mib":1159.92578125},{"t_s":198,"rss_mib":1159.93359375},{"t_s":200,"rss_mib":1159.94140625},{"t_s":201,"rss_mib":1159.94140625},{"t_s":202,"rss_mib":1159.94921875},{"t_s":204,"rss_mib":1159.94921875},{"t_s":205,"rss_mib":1159.95703125},{"t_s":207,"rss_mib":1159.95703125},{"t_s":209,"rss_mib":1159.98046875},{"t_s":210,"rss_mib":1159.98046875},{"t_s":212,"rss_mib":1159.98828125},{"t_s":213,"rss_mib":1159.98828125},{"t_s":215,"rss_mib":1160},{"t_s":216,"rss_mib":1160},{"t_s":218,"rss_mib":1160.01171875},{"t_s":219,"rss_mib":1160.01171875},{"t_s":221,"rss_mib":1160.01953125},{"t_s":222,"rss_mib":1160.01953125},{"t_s":224,"rss_mib":1160.046875},{"t_s":225,"rss_mib":1160.0546875},{"t_s":228,"rss_mib":1160.12109375},{"t_s":228,"rss_mib":1160.1328125},{"t_s":231,"rss_mib":1160.2109375},{"t_s":231,"rss_mib":1160.2109375},{"t_s":234,"rss_mib":1160.21484375},{"t_s":234,"rss_mib":1160.234375},{"t_s":236,"rss_mib":1160.2421875},{"t_s":237,"rss_mib":1160.2421875},{"t_s":240,"rss_mib":1160.2421875},{"t_s":243,"rss_mib":1160.2421875},{"t_s":246,"rss_mib":1160.2421875},{"t_s":249,"rss_mib":1160.2421875},{"t_s":250,"rss_mib":1160.25390625},{"t_s":252,"rss_mib":1160.25390625},{"t_s":254,"rss_mib":1160.26171875},{"t_s":255,"rss_mib":1160.26171875},{"t_s":256,"rss_mib":1160.26953125},{"t_s":258,"rss_mib":1160.26953125},{"t_s":261,"rss_mib":1160.26953125},{"t_s":262,"rss_mib":1160.27734375},{"t_s":264,"rss_mib":1160.27734375},{"t_s":266,"rss_mib":1160.2890625},{"t_s":267,"rss_mib":1160.2890625},{"t_s":270,"rss_mib":1160.2890625},{"t_s":273,"rss_mib":1160.2890625},{"t_s":276,"rss_mib":1160.2890625},{"t_s":278,"rss_mib":1160.3046875},{"t_s":279,"rss_mib":1160.3046875},{"t_s":281,"rss_mib":1160.30859375},{"t_s":282,"rss_mib":1160.30859375},{"t_s":285,"rss_mib":1160.33203125},{"t_s":285,"rss_mib":1160.33203125},{"t_s":288,"rss_mib":1160.44921875},{"t_s":288,"rss_mib":1160.44921875},{"t_s":291,"rss_mib":1160.44921875},{"t_s":294,"rss_mib":1160.47265625},{"t_s":294,"rss_mib":1160.48046875},{"t_s":297,"rss_mib":1160.49609375},{"t_s":297,"rss_mib":1160.49609375},{"t_s":299,"rss_mib":1160.50390625},{"t_s":300,"rss_mib":1160.50390625},{"t_s":303,"rss_mib":1160.50390625},{"t_s":306,"rss_mib":1160.50390625},{"t_s":307,"rss_mib":1160.51171875},{"t_s":309,"rss_mib":1160.51171875},{"t_s":311,"rss_mib":1160.5703125},{"t_s":312,"rss_mib":1160.5703125},{"t_s":315,"rss_mib":1160.5703125},{"t_s":318,"rss_mib":1160.5703125},{"t_s":321,"rss_mib":1160.57421875},{"t_s":321,"rss_mib":1160.57421875},{"t_s":324,"rss_mib":1160.57421875},{"t_s":325,"rss_mib":1160.58984375},{"t_s":327,"rss_mib":1160.58984375},{"t_s":329,"rss_mib":1160.59765625},{"t_s":330,"rss_mib":1160.59765625},{"t_s":332,"rss_mib":1160.60546875},{"t_s":333,"rss_mib":1160.60546875},{"t_s":336,"rss_mib":1160.60546875},{"t_s":339,"rss_mib":1160.60546875},{"t_s":342,"rss_mib":1160.60546875},{"t_s":345,"rss_mib":1160.609375},{"t_s":345,"rss_mib":1160.609375},{"t_s":346,"rss_mib":1160.69140625},{"t_s":348,"rss_mib":1160.69140625},{"t_s":351,"rss_mib":1160.6953125},{"t_s":351,"rss_mib":1160.6953125},{"t_s":354,"rss_mib":1161.40625},{"t_s":354,"rss_mib":1161.40625},{"t_s":356,"rss_mib":1161.42578125},{"t_s":357,"rss_mib":1161.42578125},{"t_s":360,"rss_mib":1161.4296875},{"t_s":360,"rss_mib":1161.4296875}],"idle_rss_series":[{"t_s":0,"rss_mib":1150.3046875},{"t_s":0,"rss_mib":1150.3984375},{"t_s":0,"rss_mib":1150.3984375},{"t_s":0,"rss_mib":1150.3984375},{"t_s":1,"rss_mib":1150.41015625},{"t_s":1,"rss_mib":1150.41015625},{"t_s":1,"rss_mib":1150.41015625},{"t_s":2,"rss_mib":1150.41015625},{"t_s":2,"rss_mib":1150.41796875},{"t_s":3,"rss_mib":1150.421875},{"t_s":3,"rss_mib":1150.421875},{"t_s":3,"rss_mib":1150.421875},{"t_s":4,"rss_mib":1150.42578125},{"t_s":4,"rss_mib":1150.4296875},{"t_s":5,"rss_mib":1150.4296875},{"t_s":5,"rss_mib":1150.4296875},{"t_s":6,"rss_mib":1150.4296875},{"t_s":6,"rss_mib":1150.4375},{"t_s":6,"rss_mib":1150.4375},{"t_s":7,"rss_mib":1150.4375},{"t_s":7,"rss_mib":1150.4375},{"t_s":8,"rss_mib":1150.4375},{"t_s":8,"rss_mib":1150.4375},{"t_s":9,"rss_mib":1150.4375},{"t_s":9,"rss_mib":1150.4453125},{"t_s":9,"rss_mib":1150.4453125},{"t_s":10,"rss_mib":1150.44921875},{"t_s":10,"rss_mib":1150.44921875},{"t_s":10,"rss_mib":1150.44921875},{"t_s":11,"rss_mib":1150.44921875},{"t_s":11,"rss_mib":1150.44921875},{"t_s":12,"rss_mib":1150.44921875},{"t_s":12,"rss_mib":1150.44921875},{"t_s":13,"rss_mib":1150.44921875},{"t_s":13,"rss_mib":1150.44921875},{"t_s":14,"rss_mib":1150.453125},{"t_s":14,"rss_mib":1150.453125},{"t_s":14,"rss_mib":1150.45703125},{"t_s":14,"rss_mib":1150.45703125},{"t_s":15,"rss_mib":1150.46484375},{"t_s":15,"rss_mib":1150.46484375},{"t_s":15,"rss_mib":1150.46484375},{"t_s":16,"rss_mib":1150.46484375},{"t_s":16,"rss_mib":1150.47265625},{"t_s":16,"rss_mib":1150.47265625},{"t_s":17,"rss_mib":1150.47265625},{"t_s":17,"rss_mib":1150.48046875},{"t_s":17,"rss_mib":1150.48046875},{"t_s":18,"rss_mib":1150.48046875},{"t_s":18,"rss_mib":1150.48046875},{"t_s":19,"rss_mib":1150.48046875},{"t_s":19,"rss_mib":1150.484375},{"t_s":19,"rss_mib":1150.484375},{"t_s":20,"rss_mib":1150.50390625},{"t_s":20,"rss_mib":1150.50390625},{"t_s":20,"rss_mib":1150.51171875},{"t_s":20,"rss_mib":1150.51171875},{"t_s":21,"rss_mib":1150.515625},{"t_s":21,"rss_mib":1150.515625},{"t_s":22,"rss_mib":1150.515625},{"t_s":22,"rss_mib":1150.515625},{"t_s":22,"rss_mib":1150.515625},{"t_s":23,"rss_mib":1150.51953125},{"t_s":23,"rss_mib":1150.5234375},{"t_s":23,"rss_mib":1150.53125},{"t_s":24,"rss_mib":1150.53125},{"t_s":24,"rss_mib":1150.54296875},{"t_s":25,"rss_mib":1150.55078125},{"t_s":25,"rss_mib":1150.55859375},{"t_s":25,"rss_mib":1150.55859375},{"t_s":25,"rss_mib":1150.55859375},{"t_s":26,"rss_mib":1150.55859375},{"t_s":26,"rss_mib":1150.55859375},{"t_s":27,"rss_mib":1150.56640625},{"t_s":27,"rss_mib":1150.5703125},{"t_s":28,"rss_mib":1150.5703125},{"t_s":28,"rss_mib":1150.5703125},{"t_s":28,"rss_mib":1150.5703125},{"t_s":29,"rss_mib":1150.578125},{"t_s":29,"rss_mib":1150.578125},{"t_s":29,"rss_mib":1150.5859375},{"t_s":29,"rss_mib":1150.5859375},{"t_s":30,"rss_mib":1150.5859375},{"t_s":30,"rss_mib":1150.59375},{"t_s":31,"rss_mib":1150.59375},{"t_s":31,"rss_mib":1150.59375},{"t_s":31,"rss_mib":1150.59375},{"t_s":32,"rss_mib":1150.59375},{"t_s":32,"rss_mib":1150.6015625},{"t_s":32,"rss_mib":1150.6015625},{"t_s":33,"rss_mib":1150.6015625},{"t_s":34,"rss_mib":1150.6015625},{"t_s":34,"rss_mib":1150.609375},{"t_s":34,"rss_mib":1150.61328125},{"t_s":34,"rss_mib":1150.62109375},{"t_s":35,"rss_mib":1150.62109375},{"t_s":35,"rss_mib":1150.62109375},{"t_s":36,"rss_mib":1150.62109375},{"t_s":36,"rss_mib":1150.62890625},{"t_s":37,"rss_mib":1150.62890625},{"t_s":37,"rss_mib":1150.62890625},{"t_s":38,"rss_mib":1150.62890625},{"t_s":38,"rss_mib":1150.62890625},{"t_s":39,"rss_mib":1150.62890625},{"t_s":39,"rss_mib":1150.62890625},{"t_s":40,"rss_mib":1150.640625},{"t_s":40,"rss_mib":1150.640625},{"t_s":41,"rss_mib":1150.640625},{"t_s":41,"rss_mib":1150.640625},{"t_s":41,"rss_mib":1150.640625},{"t_s":42,"rss_mib":1150.64453125},{"t_s":42,"rss_mib":1150.64453125},{"t_s":42,"rss_mib":1150.64453125},{"t_s":43,"rss_mib":1150.64453125},{"t_s":44,"rss_mib":1150.64453125},{"t_s":44,"rss_mib":1150.64453125},{"t_s":44,"rss_mib":1150.6484375},{"t_s":44,"rss_mib":1150.6484375},{"t_s":45,"rss_mib":1150.67578125},{"t_s":45,"rss_mib":1150.67578125},{"t_s":45,"rss_mib":1150.67578125},{"t_s":46,"rss_mib":1150.78125},{"t_s":46,"rss_mib":1150.78125},{"t_s":47,"rss_mib":1150.78125},{"t_s":47,"rss_mib":1150.78125},{"t_s":47,"rss_mib":1150.78125},{"t_s":48,"rss_mib":1150.78125},{"t_s":48,"rss_mib":1150.78125},{"t_s":49,"rss_mib":1150.78125},{"t_s":50,"rss_mib":1150.78125},{"t_s":50,"rss_mib":1150.78125},{"t_s":50,"rss_mib":1150.796875},{"t_s":50,"rss_mib":1150.796875},{"t_s":51,"rss_mib":1150.81640625},{"t_s":51,"rss_mib":1150.81640625},{"t_s":51,"rss_mib":1150.81640625},{"t_s":52,"rss_mib":1150.8203125},{"t_s":52,"rss_mib":1150.8203125},{"t_s":52,"rss_mib":1150.828125},{"t_s":53,"rss_mib":1150.828125},{"t_s":53,"rss_mib":1150.828125},{"t_s":53,"rss_mib":1150.828125},{"t_s":54,"rss_mib":1151.05078125},{"t_s":54,"rss_mib":1151.05078125},{"t_s":54,"rss_mib":1151.0625},{"t_s":54,"rss_mib":1151.0625},{"t_s":55,"rss_mib":1151.0625},{"t_s":56,"rss_mib":1151.0625},{"t_s":56,"rss_mib":1151.0703125},{"t_s":56,"rss_mib":1151.0703125},{"t_s":56,"rss_mib":1151.0703125},{"t_s":57,"rss_mib":1151.0859375},{"t_s":57,"rss_mib":1151.0859375},{"t_s":57,"rss_mib":1151.08984375},{"t_s":58,"rss_mib":1151.08984375},{"t_s":59,"rss_mib":1151.09765625},{"t_s":59,"rss_mib":1151.1015625},{"t_s":59,"rss_mib":1151.109375},{"t_s":59,"rss_mib":1151.109375}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.009739314,"cost":6.035886812,"memory":442.002045592,"streaming":0.000002958,"streams_sustained":0.000002079,"throughput":334.960594115},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"perf.frontier.1ms.concurrency":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.p99_us":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.rps":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.samples":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"stream.added_gap_p50_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_gap_p99_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p50_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p99_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_direct_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_gw_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_fps":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_headroom":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_mock_ceiling":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_direct_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_gw_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"}}},"gemini":{"served":true,"reason":null,"status":"","path":"/v1beta/models/gemini-1.5-pro:generateContent","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":750,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":932,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":959,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":27,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":737.715063745513,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1355.5369127516778,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.7426132259918599,"certified":true,"suppressed":false},"cost_window_ok":{"value":24237,"certified":true,"suppressed":false},"cost_window_rps":{"value":4038,"certified":true,"suppressed":false},"cost_threads":{"value":140,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.007014069398027809,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":5241,"certified":true,"suppressed":false},"concurrency":16,"p99_us":4564,"first_disqualified_conc":32,"lower_bound":false,"samples":31452},{"bound_ms":10,"rps":{"value":5428,"certified":true,"suppressed":false},"concurrency":32,"p99_us":8261,"first_disqualified_conc":64,"lower_bound":false,"samples":32577},{"bound_ms":50,"rps":{"value":5506,"certified":true,"suppressed":false},"concurrency":64,"p99_us":16888,"first_disqualified_conc":256,"lower_bound":false,"samples":33045},{"bound_ms":100,"rps":{"value":5506,"certified":true,"suppressed":false},"concurrency":64,"p99_us":16888,"first_disqualified_conc":512,"lower_bound":false,"samples":33045},{"bound_ms":null,"rps":{"value":5506,"certified":true,"suppressed":false},"concurrency":64,"p99_us":16888,"first_disqualified_conc":null,"lower_bound":false,"samples":33045}],"sweep":[{"conc":1,"ok":7359,"rps":1226,"p99_us":1070,"fail":0},{"conc":1,"ok":7275,"rps":1212,"p99_us":1084,"fail":0},{"conc":1,"ok":7354,"rps":1225,"p99_us":1069,"fail":0},{"conc":2,"ok":14538,"rps":2422,"p99_us":1079,"fail":0},{"conc":2,"ok":14468,"rps":2410,"p99_us":1077,"fail":0},{"conc":2,"ok":7978,"rps":1329,"p99_us":1825,"fail":0},{"conc":4,"ok":8004,"rps":1333,"p99_us":3801,"fail":0},{"conc":4,"ok":22330,"rps":3721,"p99_us":1838,"fail":0},{"conc":4,"ok":22463,"rps":3743,"p99_us":1839,"fail":0},{"conc":8,"ok":23127,"rps":3854,"p99_us":3551,"fail":0},{"conc":8,"ok":29800,"rps":4965,"p99_us":2931,"fail":0},{"conc":8,"ok":30454,"rps":5074,"p99_us":2797,"fail":0},{"conc":16,"ok":31533,"rps":5255,"p99_us":5327,"fail":0},{"conc":16,"ok":31441,"rps":5239,"p99_us":5447,"fail":0},{"conc":16,"ok":31452,"rps":5241,"p99_us":4564,"fail":0},{"conc":32,"ok":32355,"rps":5391,"p99_us":10917,"fail":0},{"conc":32,"ok":32366,"rps":5394,"p99_us":11069,"fail":0},{"conc":32,"ok":32577,"rps":5428,"p99_us":8261,"fail":0},{"conc":64,"ok":33045,"rps":5506,"p99_us":16888,"fail":0},{"conc":64,"ok":32866,"rps":5476,"p99_us":17037,"fail":0},{"conc":64,"ok":32252,"rps":5375,"p99_us":17624,"fail":0},{"conc":128,"ok":31934,"rps":5320,"p99_us":35075,"fail":0},{"conc":128,"ok":31954,"rps":5324,"p99_us":31605,"fail":0},{"conc":128,"ok":31963,"rps":5325,"p99_us":32926,"fail":0},{"conc":256,"ok":31044,"rps":5173,"p99_us":63565,"fail":0},{"conc":256,"ok":31479,"rps":5245,"p99_us":63646,"fail":0},{"conc":256,"ok":31622,"rps":5269,"p99_us":66456,"fail":0},{"conc":512,"ok":30528,"rps":5087,"p99_us":135578,"fail":0},{"conc":512,"ok":30609,"rps":5101,"p99_us":125811,"fail":0},{"conc":512,"ok":30640,"rps":5105,"p99_us":129816,"fail":0},{"conc":1024,"ok":30621,"rps":5102,"p99_us":1535458,"fail":0},{"conc":1024,"ok":30859,"rps":5142,"p99_us":279600,"fail":0},{"conc":1024,"ok":30780,"rps":5129,"p99_us":290850,"fail":0},{"conc":2048,"ok":30364,"rps":5059,"p99_us":879266,"fail":0},{"conc":2048,"ok":30566,"rps":5093,"p99_us":780885,"fail":0},{"conc":2048,"ok":30703,"rps":5116,"p99_us":839865,"fail":0},{"conc":4096,"ok":30606,"rps":5099,"p99_us":875234,"fail":0},{"conc":4096,"ok":30175,"rps":5028,"p99_us":859319,"fail":0},{"conc":4096,"ok":30404,"rps":5066,"p99_us":856930,"fail":0},{"conc":8192,"ok":30514,"rps":5084,"p99_us":914306,"fail":0},{"conc":8192,"ok":30172,"rps":5027,"p99_us":918328,"fail":0},{"conc":8192,"ok":30340,"rps":5055,"p99_us":967613,"fail":0},{"conc":16384,"ok":30532,"rps":5088,"p99_us":877399,"fail":0},{"conc":16384,"ok":30380,"rps":5062,"p99_us":2265548,"fail":0},{"conc":16384,"ok":30312,"rps":5051,"p99_us":443705,"fail":0},{"conc":32768,"ok":27141,"rps":4522,"p99_us":3076166,"fail":0},{"conc":32768,"ok":25497,"rps":4248,"p99_us":2757305,"fail":0},{"conc":32768,"ok":27250,"rps":4540,"p99_us":2640518,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1225,"certified":true,"suppressed":false},"p99_us":1070,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":2410,"certified":true,"suppressed":false},"p99_us":1079,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":3721,"certified":true,"suppressed":false},"p99_us":1839,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":4965,"certified":true,"suppressed":false},"p99_us":2931,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":5241,"certified":true,"suppressed":false},"p99_us":5327,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":5394,"certified":true,"suppressed":false},"p99_us":10917,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":5476,"certified":true,"suppressed":false},"p99_us":17037,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":5324,"certified":true,"suppressed":false},"p99_us":32926,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":5245,"certified":true,"suppressed":false},"p99_us":63646,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":5101,"certified":true,"suppressed":false},"p99_us":129816,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":5129,"certified":true,"suppressed":false},"p99_us":290850,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":5093,"certified":true,"suppressed":false},"p99_us":839865,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":5066,"certified":true,"suppressed":false},"p99_us":859319,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":5055,"certified":true,"suppressed":false},"p99_us":918328,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":5062,"certified":true,"suppressed":false},"p99_us":877399,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":4522,"certified":true,"suppressed":false},"p99_us":2757305,"windows":3,"clean_windows":3}],"reverify_note":"gemini>gemini is a same-dialect cell: there is no translation to prove, so a request-shape check upstream cannot distinguish a translating gateway from a forwarding one","c1_note":"the c=1 percentiles are taken over 7714 successful gateway round trip(s) and 265544 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":7714,"direct_c1_samples":265544},"stream":{"stream_served":"untestable","stream_error":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","reason":"untestable","stream_c1_note":null,"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"}},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1156.921875,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1165.796875,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1166.64453125,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1175.75390625,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1269.25,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":66,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":1.08419810667365,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":0.8945988670581346,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1157.5},{"t_s":2,"rss_mib":1174.1328125},{"t_s":3,"rss_mib":1174.140625},{"t_s":4,"rss_mib":1174.17578125},{"t_s":6,"rss_mib":1174.17578125},{"t_s":8,"rss_mib":1174.4609375},{"t_s":9,"rss_mib":1174.4609375},{"t_s":11,"rss_mib":1174.50390625},{"t_s":12,"rss_mib":1174.59375},{"t_s":14,"rss_mib":1174.64453125},{"t_s":15,"rss_mib":1174.64453125},{"t_s":17,"rss_mib":1174.67578125},{"t_s":18,"rss_mib":1174.734375},{"t_s":20,"rss_mib":1174.77734375},{"t_s":21,"rss_mib":1174.77734375},{"t_s":23,"rss_mib":1174.8125},{"t_s":24,"rss_mib":1174.8125},{"t_s":25,"rss_mib":1174.8359375},{"t_s":27,"rss_mib":1174.8359375},{"t_s":29,"rss_mib":1174.8828125},{"t_s":30,"rss_mib":1174.90234375},{"t_s":32,"rss_mib":1175.015625},{"t_s":33,"rss_mib":1175.0234375},{"t_s":34,"rss_mib":1175.05078125},{"t_s":36,"rss_mib":1175.05078125},{"t_s":36,"rss_mib":1175.07421875},{"t_s":39,"rss_mib":1175.0859375},{"t_s":41,"rss_mib":1175.1171875},{"t_s":42,"rss_mib":1175.15625},{"t_s":44,"rss_mib":1175.34765625},{"t_s":45,"rss_mib":1175.34765625},{"t_s":47,"rss_mib":1175.3828125},{"t_s":48,"rss_mib":1175.39453125},{"t_s":49,"rss_mib":1175.609375},{"t_s":51,"rss_mib":1175.609375},{"t_s":53,"rss_mib":1175.63671875},{"t_s":54,"rss_mib":1175.6875},{"t_s":55,"rss_mib":1175.74609375},{"t_s":59,"rss_mib":1175.75390625},{"t_s":59,"rss_mib":1173.5703125},{"t_s":60,"rss_mib":1173.5703125},{"t_s":62,"rss_mib":1173.5859375},{"t_s":63,"rss_mib":1173.62109375},{"t_s":66,"rss_mib":1173.6328125},{"t_s":69,"rss_mib":1173.6328125},{"t_s":72,"rss_mib":1173.6328125},{"t_s":72,"rss_mib":1173.63671875},{"t_s":75,"rss_mib":1173.63671875},{"t_s":76,"rss_mib":1173.64453125},{"t_s":78,"rss_mib":1173.64453125},{"t_s":79,"rss_mib":1173.6484375},{"t_s":81,"rss_mib":1173.6484375},{"t_s":84,"rss_mib":1173.66015625},{"t_s":86,"rss_mib":1173.71484375},{"t_s":87,"rss_mib":1173.71484375},{"t_s":89,"rss_mib":1173.7265625},{"t_s":90,"rss_mib":1173.7421875},{"t_s":92,"rss_mib":1173.796875},{"t_s":93,"rss_mib":1173.796875},{"t_s":94,"rss_mib":1173.8125},{"t_s":96,"rss_mib":1173.875},{"t_s":98,"rss_mib":1173.890625},{"t_s":99,"rss_mib":1173.890625},{"t_s":101,"rss_mib":1173.9140625},{"t_s":102,"rss_mib":1173.9140625},{"t_s":104,"rss_mib":1173.96875},{"t_s":105,"rss_mib":1173.96875},{"t_s":107,"rss_mib":1173.9765625},{"t_s":108,"rss_mib":1173.9765625},{"t_s":111,"rss_mib":1173.9921875},{"t_s":111,"rss_mib":1173.9921875},{"t_s":111,"rss_mib":1174},{"t_s":114,"rss_mib":1174.0078125},{"t_s":115,"rss_mib":1174.09765625},{"t_s":117,"rss_mib":1174.09765625},{"t_s":119,"rss_mib":1174.125},{"t_s":120,"rss_mib":1174.125},{"t_s":121,"rss_mib":1174.1328125},{"t_s":123,"rss_mib":1174.1328125},{"t_s":123,"rss_mib":1174.16015625},{"t_s":126,"rss_mib":1174.16015625},{"t_s":127,"rss_mib":1174.17578125},{"t_s":129,"rss_mib":1174.17578125},{"t_s":130,"rss_mib":1174.18359375},{"t_s":132,"rss_mib":1174.18359375},{"t_s":135,"rss_mib":1174.234375},{"t_s":135,"rss_mib":1174.234375},{"t_s":136,"rss_mib":1174.27734375},{"t_s":138,"rss_mib":1174.27734375},{"t_s":141,"rss_mib":1174.296875},{"t_s":141,"rss_mib":1174.296875},{"t_s":143,"rss_mib":1174.3125},{"t_s":144,"rss_mib":1174.3125},{"t_s":146,"rss_mib":1174.328125},{"t_s":147,"rss_mib":1174.328125},{"t_s":148,"rss_mib":1174.34375},{"t_s":150,"rss_mib":1174.34375},{"t_s":152,"rss_mib":1174.359375},{"t_s":153,"rss_mib":1174.359375},{"t_s":155,"rss_mib":1174.3828125},{"t_s":156,"rss_mib":1174.390625},{"t_s":159,"rss_mib":1174.40625},{"t_s":159,"rss_mib":1174.40625},{"t_s":161,"rss_mib":1174.4140625},{"t_s":162,"rss_mib":1174.4140625},{"t_s":164,"rss_mib":1174.61328125},{"t_s":165,"rss_mib":1174.61328125},{"t_s":167,"rss_mib":1174.63671875},{"t_s":168,"rss_mib":1174.63671875},{"t_s":170,"rss_mib":1169.203125},{"t_s":171,"rss_mib":1169.203125},{"t_s":173,"rss_mib":1169.20703125},{"t_s":174,"rss_mib":1167.1796875},{"t_s":175,"rss_mib":1167.19140625},{"t_s":177,"rss_mib":1167.19140625},{"t_s":179,"rss_mib":1167.21484375},{"t_s":180,"rss_mib":1167.21484375},{"t_s":180,"rss_mib":1167.22265625},{"t_s":183,"rss_mib":1167.22265625},{"t_s":184,"rss_mib":1166.58984375},{"t_s":186,"rss_mib":1166.59765625},{"t_s":188,"rss_mib":1166.640625},{"t_s":189,"rss_mib":1166.640625},{"t_s":189,"rss_mib":1164.9296875},{"t_s":192,"rss_mib":1165.01171875},{"t_s":193,"rss_mib":1165.09765625},{"t_s":195,"rss_mib":1165.09765625},{"t_s":197,"rss_mib":1165.11328125},{"t_s":198,"rss_mib":1165.11328125},{"t_s":201,"rss_mib":1165.11328125},{"t_s":204,"rss_mib":1165.125},{"t_s":204,"rss_mib":1165.125},{"t_s":204,"rss_mib":1165.1328125},{"t_s":207,"rss_mib":1165.1328125},{"t_s":209,"rss_mib":1165.16015625},{"t_s":210,"rss_mib":1165.16015625},{"t_s":211,"rss_mib":1165.16796875},{"t_s":213,"rss_mib":1165.16796875},{"t_s":215,"rss_mib":1165.1796875},{"t_s":216,"rss_mib":1165.1796875},{"t_s":218,"rss_mib":1165.1875},{"t_s":219,"rss_mib":1165.1875},{"t_s":221,"rss_mib":1165.1953125},{"t_s":222,"rss_mib":1165.1953125},{"t_s":224,"rss_mib":1165.25390625},{"t_s":225,"rss_mib":1165.25390625},{"t_s":227,"rss_mib":1165.26171875},{"t_s":228,"rss_mib":1165.2734375},{"t_s":231,"rss_mib":1165.296875},{"t_s":231,"rss_mib":1165.296875},{"t_s":234,"rss_mib":1165.40625},{"t_s":234,"rss_mib":1165.40625},{"t_s":235,"rss_mib":1165.4296875},{"t_s":237,"rss_mib":1165.4296875},{"t_s":240,"rss_mib":1165.4453125},{"t_s":240,"rss_mib":1165.4453125},{"t_s":243,"rss_mib":1165.453125},{"t_s":243,"rss_mib":1165.453125},{"t_s":243,"rss_mib":1165.4609375},{"t_s":246,"rss_mib":1165.4609375},{"t_s":249,"rss_mib":1165.4609375},{"t_s":251,"rss_mib":1165.67578125},{"t_s":252,"rss_mib":1165.6953125},{"t_s":254,"rss_mib":1165.76171875},{"t_s":255,"rss_mib":1165.76171875},{"t_s":256,"rss_mib":1165.78515625},{"t_s":258,"rss_mib":1165.78515625},{"t_s":259,"rss_mib":1165.79296875},{"t_s":261,"rss_mib":1165.79296875},{"t_s":262,"rss_mib":1165.796875},{"t_s":264,"rss_mib":1165.796875},{"t_s":267,"rss_mib":1165.796875},{"t_s":270,"rss_mib":1165.80859375},{"t_s":272,"rss_mib":1165.8125},{"t_s":273,"rss_mib":1165.8125},{"t_s":274,"rss_mib":1165.81640625},{"t_s":276,"rss_mib":1165.81640625},{"t_s":277,"rss_mib":1165.96875},{"t_s":279,"rss_mib":1165.96875},{"t_s":281,"rss_mib":1166},{"t_s":282,"rss_mib":1166},{"t_s":283,"rss_mib":1166.1796875},{"t_s":285,"rss_mib":1166.1796875},{"t_s":287,"rss_mib":1166.328125},{"t_s":288,"rss_mib":1166.328125},{"t_s":289,"rss_mib":1166.39453125},{"t_s":294,"rss_mib":1166.36328125},{"t_s":294,"rss_mib":1166.46875},{"t_s":294,"rss_mib":1166.46875},{"t_s":295,"rss_mib":1166.47265625},{"t_s":297,"rss_mib":1166.47265625},{"t_s":299,"rss_mib":1166.48046875},{"t_s":300,"rss_mib":1166.48046875},{"t_s":302,"rss_mib":1166.484375},{"t_s":303,"rss_mib":1166.484375},{"t_s":306,"rss_mib":1166.484375},{"t_s":309,"rss_mib":1166.484375},{"t_s":312,"rss_mib":1166.484375},{"t_s":313,"rss_mib":1166.48828125},{"t_s":315,"rss_mib":1166.48828125},{"t_s":317,"rss_mib":1166.5},{"t_s":318,"rss_mib":1166.5078125},{"t_s":321,"rss_mib":1166.5078125},{"t_s":324,"rss_mib":1166.5078125},{"t_s":327,"rss_mib":1166.5078125},{"t_s":330,"rss_mib":1166.5078125},{"t_s":333,"rss_mib":1166.5078125},{"t_s":336,"rss_mib":1166.515625},{"t_s":336,"rss_mib":1166.515625},{"t_s":337,"rss_mib":1166.53515625},{"t_s":339,"rss_mib":1166.53515625},{"t_s":341,"rss_mib":1166.56640625},{"t_s":342,"rss_mib":1166.56640625},{"t_s":343,"rss_mib":1166.64453125},{"t_s":345,"rss_mib":1166.64453125},{"t_s":347,"rss_mib":1166.6640625},{"t_s":348,"rss_mib":1166.6640625},{"t_s":350,"rss_mib":1166.6796875},{"t_s":351,"rss_mib":1166.6796875},{"t_s":354,"rss_mib":1166.76171875},{"t_s":354,"rss_mib":1166.76953125},{"t_s":356,"rss_mib":1166.77734375},{"t_s":357,"rss_mib":1166.77734375},{"t_s":360,"rss_mib":1166.796875},{"t_s":360,"rss_mib":1166.796875}],"idle_rss_series":[{"t_s":0,"rss_mib":1156.3671875},{"t_s":0,"rss_mib":1156.4296875},{"t_s":0,"rss_mib":1156.4296875},{"t_s":0,"rss_mib":1156.4453125},{"t_s":0,"rss_mib":1156.4453125},{"t_s":1,"rss_mib":1156.4453125},{"t_s":1,"rss_mib":1156.453125},{"t_s":1,"rss_mib":1156.45703125},{"t_s":2,"rss_mib":1156.45703125},{"t_s":2,"rss_mib":1156.5703125},{"t_s":2,"rss_mib":1156.5703125},{"t_s":3,"rss_mib":1156.578125},{"t_s":3,"rss_mib":1156.5859375},{"t_s":3,"rss_mib":1156.5859375},{"t_s":4,"rss_mib":1156.59765625},{"t_s":4,"rss_mib":1156.59765625},{"t_s":4,"rss_mib":1156.59765625},{"t_s":5,"rss_mib":1156.59765625},{"t_s":5,"rss_mib":1156.61328125},{"t_s":5,"rss_mib":1156.61328125},{"t_s":6,"rss_mib":1156.61328125},{"t_s":6,"rss_mib":1156.61328125},{"t_s":7,"rss_mib":1156.62109375},{"t_s":7,"rss_mib":1156.62109375},{"t_s":8,"rss_mib":1156.62890625},{"t_s":8,"rss_mib":1156.62890625},{"t_s":8,"rss_mib":1156.62890625},{"t_s":9,"rss_mib":1156.63671875},{"t_s":9,"rss_mib":1156.63671875},{"t_s":9,"rss_mib":1156.6484375},{"t_s":9,"rss_mib":1156.65234375},{"t_s":10,"rss_mib":1156.66015625},{"t_s":10,"rss_mib":1156.66015625},{"t_s":10,"rss_mib":1156.6640625},{"t_s":11,"rss_mib":1156.671875},{"t_s":11,"rss_mib":1156.671875},{"t_s":11,"rss_mib":1156.671875},{"t_s":12,"rss_mib":1156.6796875},{"t_s":12,"rss_mib":1156.6875},{"t_s":12,"rss_mib":1156.69140625},{"t_s":13,"rss_mib":1156.69140625},{"t_s":13,"rss_mib":1156.7109375},{"t_s":14,"rss_mib":1156.71875},{"t_s":14,"rss_mib":1156.71875},{"t_s":14,"rss_mib":1156.7265625},{"t_s":14,"rss_mib":1156.7265625},{"t_s":15,"rss_mib":1156.734375},{"t_s":15,"rss_mib":1156.734375},{"t_s":15,"rss_mib":1156.7421875},{"t_s":16,"rss_mib":1156.75},{"t_s":16,"rss_mib":1156.75},{"t_s":16,"rss_mib":1156.7578125},{"t_s":16,"rss_mib":1156.7578125},{"t_s":17,"rss_mib":1156.7578125},{"t_s":17,"rss_mib":1156.7578125},{"t_s":18,"rss_mib":1156.765625},{"t_s":18,"rss_mib":1156.765625},{"t_s":18,"rss_mib":1156.765625},{"t_s":19,"rss_mib":1156.7734375},{"t_s":19,"rss_mib":1156.7734375},{"t_s":19,"rss_mib":1156.78125},{"t_s":19,"rss_mib":1156.78125},{"t_s":20,"rss_mib":1156.78125},{"t_s":20,"rss_mib":1156.7890625},{"t_s":20,"rss_mib":1156.7890625},{"t_s":21,"rss_mib":1156.796875},{"t_s":21,"rss_mib":1156.8046875},{"t_s":22,"rss_mib":1156.8046875},{"t_s":22,"rss_mib":1156.8046875},{"t_s":22,"rss_mib":1156.8046875},{"t_s":23,"rss_mib":1156.8125},{"t_s":23,"rss_mib":1156.8125},{"t_s":23,"rss_mib":1156.8125},{"t_s":24,"rss_mib":1156.83984375},{"t_s":24,"rss_mib":1156.84375},{"t_s":24,"rss_mib":1156.8515625},{"t_s":24,"rss_mib":1156.85546875},{"t_s":25,"rss_mib":1156.8671875},{"t_s":25,"rss_mib":1156.8671875},{"t_s":25,"rss_mib":1156.875},{"t_s":26,"rss_mib":1156.875},{"t_s":26,"rss_mib":1156.8828125},{"t_s":26,"rss_mib":1156.890625},{"t_s":27,"rss_mib":1156.8984375},{"t_s":27,"rss_mib":1156.8984375},{"t_s":27,"rss_mib":1156.8984375},{"t_s":28,"rss_mib":1156.8984375},{"t_s":28,"rss_mib":1156.90625},{"t_s":29,"rss_mib":1156.9140625},{"t_s":29,"rss_mib":1156.9140625},{"t_s":29,"rss_mib":1156.921875},{"t_s":29,"rss_mib":1156.921875},{"t_s":30,"rss_mib":1156.921875},{"t_s":30,"rss_mib":1156.92578125},{"t_s":31,"rss_mib":1156.92578125},{"t_s":31,"rss_mib":1156.93359375},{"t_s":31,"rss_mib":1156.94140625},{"t_s":32,"rss_mib":1156.94921875},{"t_s":32,"rss_mib":1156.94921875},{"t_s":32,"rss_mib":1156.94921875},{"t_s":33,"rss_mib":1156.94921875},{"t_s":33,"rss_mib":1156.96484375},{"t_s":33,"rss_mib":1156.96875},{"t_s":34,"rss_mib":1156.96875},{"t_s":34,"rss_mib":1156.9765625},{"t_s":34,"rss_mib":1156.9765625},{"t_s":35,"rss_mib":1156.9765625},{"t_s":35,"rss_mib":1156.9765625},{"t_s":36,"rss_mib":1156.984375},{"t_s":36,"rss_mib":1156.9921875},{"t_s":36,"rss_mib":1157},{"t_s":36,"rss_mib":1157},{"t_s":37,"rss_mib":1157},{"t_s":37,"rss_mib":1157.0078125},{"t_s":38,"rss_mib":1157.01171875},{"t_s":38,"rss_mib":1157.01171875},{"t_s":38,"rss_mib":1157.01953125},{"t_s":38,"rss_mib":1157.01953125},{"t_s":39,"rss_mib":1157.01953125},{"t_s":39,"rss_mib":1157.02734375},{"t_s":39,"rss_mib":1157.03515625},{"t_s":40,"rss_mib":1157.03515625},{"t_s":40,"rss_mib":1157.04296875},{"t_s":41,"rss_mib":1157.04296875},{"t_s":41,"rss_mib":1157.04296875},{"t_s":41,"rss_mib":1157.05078125},{"t_s":41,"rss_mib":1157.05078125},{"t_s":42,"rss_mib":1157.05078125},{"t_s":42,"rss_mib":1157.0546875},{"t_s":42,"rss_mib":1157.0546875},{"t_s":43,"rss_mib":1157.0625},{"t_s":43,"rss_mib":1157.0703125},{"t_s":43,"rss_mib":1157.125},{"t_s":44,"rss_mib":1157.125},{"t_s":44,"rss_mib":1157.125},{"t_s":44,"rss_mib":1157.1328125},{"t_s":44,"rss_mib":1157.1328125},{"t_s":45,"rss_mib":1157.1328125},{"t_s":45,"rss_mib":1157.1484375},{"t_s":45,"rss_mib":1157.1484375},{"t_s":46,"rss_mib":1157.15625},{"t_s":46,"rss_mib":1157.1640625},{"t_s":47,"rss_mib":1157.171875},{"t_s":47,"rss_mib":1157.171875},{"t_s":47,"rss_mib":1157.1796875},{"t_s":47,"rss_mib":1157.1796875},{"t_s":48,"rss_mib":1157.1796875},{"t_s":48,"rss_mib":1157.1796875},{"t_s":49,"rss_mib":1157.1953125},{"t_s":49,"rss_mib":1157.1953125},{"t_s":49,"rss_mib":1157.203125},{"t_s":50,"rss_mib":1157.203125},{"t_s":50,"rss_mib":1157.203125},{"t_s":51,"rss_mib":1157.203125},{"t_s":51,"rss_mib":1157.21875},{"t_s":51,"rss_mib":1157.21875},{"t_s":52,"rss_mib":1157.2265625},{"t_s":52,"rss_mib":1157.2265625},{"t_s":52,"rss_mib":1157.234375},{"t_s":52,"rss_mib":1157.234375},{"t_s":53,"rss_mib":1157.2578125},{"t_s":53,"rss_mib":1157.2578125},{"t_s":53,"rss_mib":1157.26953125},{"t_s":54,"rss_mib":1157.3828125},{"t_s":54,"rss_mib":1157.3828125},{"t_s":54,"rss_mib":1157.39453125},{"t_s":54,"rss_mib":1157.39453125},{"t_s":55,"rss_mib":1157.39453125},{"t_s":55,"rss_mib":1157.39453125},{"t_s":56,"rss_mib":1157.40234375},{"t_s":56,"rss_mib":1157.40234375},{"t_s":56,"rss_mib":1157.41015625},{"t_s":56,"rss_mib":1157.41015625},{"t_s":57,"rss_mib":1157.421875},{"t_s":57,"rss_mib":1157.421875},{"t_s":57,"rss_mib":1157.421875},{"t_s":58,"rss_mib":1157.4296875},{"t_s":58,"rss_mib":1157.4296875},{"t_s":58,"rss_mib":1157.4453125},{"t_s":59,"rss_mib":1157.4453125},{"t_s":59,"rss_mib":1157.4453125}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.009577532,"cost":6.034147351,"memory":433.915940866,"streaming":0.000002593,"streams_sustained":0.000001552,"throughput":327.226214967},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"perf.frontier.1ms.concurrency":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.p99_us":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.rps":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.samples":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"stream.added_gap_p50_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_gap_p99_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p50_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p99_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_direct_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_gw_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_fps":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_headroom":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_mock_ceiling":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_direct_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_gw_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"}}},"bedrock":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/model/gemini-1.5-pro/converse","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"cohere":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v2/chat","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"openai-responses":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1/responses","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"anthropic":{"served":true,"reason":null,"status":"","path":"/v1/messages","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":1045,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":1360,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":1388,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":1055.3296800424253,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":947.571189279732,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.7438065009330738,"certified":true,"suppressed":false},"cost_window_ok":{"value":16971,"certified":true,"suppressed":false},"cost_window_rps":{"value":2828,"certified":true,"suppressed":false},"cost_threads":{"value":142,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.01019386011431265,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":3821,"certified":true,"suppressed":false},"concurrency":8,"p99_us":2813,"first_disqualified_conc":16,"lower_bound":false,"samples":22929},{"bound_ms":10,"rps":{"value":3862,"certified":true,"suppressed":false},"concurrency":16,"p99_us":8825,"first_disqualified_conc":32,"lower_bound":false,"samples":23185},{"bound_ms":50,"rps":{"value":3961,"certified":true,"suppressed":false},"concurrency":64,"p99_us":26997,"first_disqualified_conc":256,"lower_bound":false,"samples":23771},{"bound_ms":100,"rps":{"value":3961,"certified":true,"suppressed":false},"concurrency":64,"p99_us":26997,"first_disqualified_conc":512,"lower_bound":false,"samples":23771},{"bound_ms":null,"rps":{"value":3961,"certified":true,"suppressed":false},"concurrency":64,"p99_us":26997,"first_disqualified_conc":null,"lower_bound":false,"samples":23771}],"sweep":[{"conc":1,"ok":5479,"rps":912,"p99_us":1409,"fail":0},{"conc":1,"ok":5419,"rps":902,"p99_us":1428,"fail":0},{"conc":1,"ok":5492,"rps":915,"p99_us":1396,"fail":0},{"conc":2,"ok":9903,"rps":1650,"p99_us":1494,"fail":0},{"conc":2,"ok":5860,"rps":975,"p99_us":2429,"fail":0},{"conc":2,"ok":10827,"rps":1804,"p99_us":1406,"fail":0},{"conc":4,"ok":16694,"rps":2781,"p99_us":2586,"fail":0},{"conc":4,"ok":16735,"rps":2788,"p99_us":2580,"fail":0},{"conc":4,"ok":16680,"rps":2779,"p99_us":2411,"fail":0},{"conc":8,"ok":22929,"rps":3821,"p99_us":2813,"fail":0},{"conc":8,"ok":22584,"rps":3763,"p99_us":3722,"fail":0},{"conc":8,"ok":17389,"rps":2897,"p99_us":4740,"fail":0},{"conc":16,"ok":23185,"rps":3862,"p99_us":8825,"fail":0},{"conc":16,"ok":17659,"rps":2942,"p99_us":8954,"fail":0},{"conc":16,"ok":22907,"rps":3817,"p99_us":7451,"fail":0},{"conc":32,"ok":23565,"rps":3927,"p99_us":12950,"fail":0},{"conc":32,"ok":23567,"rps":3927,"p99_us":13975,"fail":0},{"conc":32,"ok":23621,"rps":3936,"p99_us":16026,"fail":0},{"conc":64,"ok":23771,"rps":3961,"p99_us":26997,"fail":0},{"conc":64,"ok":23758,"rps":3958,"p99_us":23457,"fail":0},{"conc":64,"ok":23756,"rps":3958,"p99_us":22701,"fail":0},{"conc":128,"ok":23634,"rps":3937,"p99_us":45363,"fail":0},{"conc":128,"ok":23658,"rps":3942,"p99_us":43353,"fail":0},{"conc":128,"ok":23487,"rps":3913,"p99_us":48518,"fail":0},{"conc":256,"ok":22804,"rps":3799,"p99_us":85875,"fail":0},{"conc":256,"ok":22843,"rps":3806,"p99_us":85712,"fail":0},{"conc":256,"ok":22875,"rps":3811,"p99_us":88357,"fail":0},{"conc":512,"ok":22472,"rps":3744,"p99_us":176172,"fail":0},{"conc":512,"ok":22414,"rps":3734,"p99_us":189484,"fail":0},{"conc":512,"ok":22466,"rps":3744,"p99_us":169667,"fail":0},{"conc":1024,"ok":22414,"rps":3734,"p99_us":2242687,"fail":0},{"conc":1024,"ok":22382,"rps":3729,"p99_us":1342415,"fail":0},{"conc":1024,"ok":22454,"rps":3741,"p99_us":993612,"fail":0},{"conc":2048,"ok":21917,"rps":3652,"p99_us":1197623,"fail":0},{"conc":2048,"ok":22095,"rps":3681,"p99_us":1264181,"fail":0},{"conc":2048,"ok":22032,"rps":3671,"p99_us":1215380,"fail":0},{"conc":4096,"ok":22089,"rps":3680,"p99_us":1227477,"fail":0},{"conc":4096,"ok":22106,"rps":3683,"p99_us":1131536,"fail":0},{"conc":4096,"ok":21874,"rps":3645,"p99_us":1155973,"fail":0},{"conc":8192,"ok":22157,"rps":3691,"p99_us":1045082,"fail":0},{"conc":8192,"ok":22048,"rps":3673,"p99_us":1085054,"fail":0},{"conc":8192,"ok":22147,"rps":3690,"p99_us":1133771,"fail":0},{"conc":16384,"ok":22165,"rps":3693,"p99_us":1072402,"fail":0},{"conc":16384,"ok":22054,"rps":3675,"p99_us":1112988,"fail":0},{"conc":16384,"ok":22224,"rps":3703,"p99_us":874572,"fail":0},{"conc":32768,"ok":17247,"rps":2874,"p99_us":2897758,"fail":0},{"conc":32768,"ok":19905,"rps":3316,"p99_us":2607174,"fail":0},{"conc":32768,"ok":19678,"rps":3278,"p99_us":2605463,"fail":0}],"rungs":[{"conc":1,"rps":{"value":912,"certified":true,"suppressed":false},"p99_us":1409,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":1650,"certified":true,"suppressed":false},"p99_us":1494,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":2781,"certified":true,"suppressed":false},"p99_us":2580,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":3763,"certified":true,"suppressed":false},"p99_us":3722,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":3817,"certified":true,"suppressed":false},"p99_us":8825,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":3927,"certified":true,"suppressed":false},"p99_us":13975,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":3958,"certified":true,"suppressed":false},"p99_us":23457,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":3937,"certified":true,"suppressed":false},"p99_us":45363,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":3806,"certified":true,"suppressed":false},"p99_us":85875,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":3744,"certified":true,"suppressed":false},"p99_us":176172,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":3734,"certified":true,"suppressed":false},"p99_us":1342415,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":3671,"certified":true,"suppressed":false},"p99_us":1215380,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":3680,"certified":true,"suppressed":false},"p99_us":1155973,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":3690,"certified":true,"suppressed":false},"p99_us":1085054,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":3693,"certified":true,"suppressed":false},"p99_us":1072402,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":3278,"certified":true,"suppressed":false},"p99_us":2607174,"windows":3,"clean_windows":3}],"egress_reverified":true,"c1_note":"the c=1 percentiles are taken over 5488 successful gateway round trip(s) and 262925 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":5488,"direct_c1_samples":262925},"stream":{"stream_served":"untestable","stream_error":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","reason":"untestable","stream_c1_note":null,"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"}},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1151.53125,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1164.61328125,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1165.828125,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1173.3046875,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1280.23828125,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":1.1668804149685401,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":1.20687057093733,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1152.36328125},{"t_s":2,"rss_mib":1170.609375},{"t_s":3,"rss_mib":1170.609375},{"t_s":5,"rss_mib":1171.12109375},{"t_s":6,"rss_mib":1171.12109375},{"t_s":8,"rss_mib":1171.28515625},{"t_s":9,"rss_mib":1171.28515625},{"t_s":11,"rss_mib":1171.34765625},{"t_s":12,"rss_mib":1171.34765625},{"t_s":14,"rss_mib":1171.54296875},{"t_s":15,"rss_mib":1171.54296875},{"t_s":17,"rss_mib":1171.58203125},{"t_s":18,"rss_mib":1171.58203125},{"t_s":19,"rss_mib":1171.91796875},{"t_s":21,"rss_mib":1171.91796875},{"t_s":23,"rss_mib":1171.94921875},{"t_s":24,"rss_mib":1171.95703125},{"t_s":25,"rss_mib":1171.98046875},{"t_s":27,"rss_mib":1171.98046875},{"t_s":28,"rss_mib":1172.01171875},{"t_s":30,"rss_mib":1172.01953125},{"t_s":32,"rss_mib":1172.05859375},{"t_s":33,"rss_mib":1172.05859375},{"t_s":34,"rss_mib":1172.09375},{"t_s":36,"rss_mib":1172.09375},{"t_s":38,"rss_mib":1172.1328125},{"t_s":39,"rss_mib":1172.1328125},{"t_s":41,"rss_mib":1172.1796875},{"t_s":42,"rss_mib":1172.18359375},{"t_s":44,"rss_mib":1172.3203125},{"t_s":45,"rss_mib":1172.3203125},{"t_s":46,"rss_mib":1172.34375},{"t_s":48,"rss_mib":1172.34375},{"t_s":49,"rss_mib":1172.5234375},{"t_s":51,"rss_mib":1172.5234375},{"t_s":54,"rss_mib":1172.5546875},{"t_s":54,"rss_mib":1172.15625},{"t_s":57,"rss_mib":1172.23046875},{"t_s":59,"rss_mib":1172.34765625},{"t_s":60,"rss_mib":1172.3515625},{"t_s":62,"rss_mib":1172.35546875},{"t_s":63,"rss_mib":1172.359375},{"t_s":63,"rss_mib":1172.390625},{"t_s":66,"rss_mib":1172.390625},{"t_s":68,"rss_mib":1172.39453125},{"t_s":69,"rss_mib":1172.39453125},{"t_s":72,"rss_mib":1172.39453125},{"t_s":75,"rss_mib":1172.39453125},{"t_s":78,"rss_mib":1172.39453125},{"t_s":81,"rss_mib":1172.39453125},{"t_s":82,"rss_mib":1172.4140625},{"t_s":84,"rss_mib":1172.4140625},{"t_s":86,"rss_mib":1172.4453125},{"t_s":87,"rss_mib":1172.4453125},{"t_s":90,"rss_mib":1172.46875},{"t_s":90,"rss_mib":1172.46875},{"t_s":92,"rss_mib":1172.48828125},{"t_s":93,"rss_mib":1172.48828125},{"t_s":95,"rss_mib":1172.51171875},{"t_s":96,"rss_mib":1172.51171875},{"t_s":98,"rss_mib":1172.52734375},{"t_s":99,"rss_mib":1172.52734375},{"t_s":101,"rss_mib":1172.55078125},{"t_s":102,"rss_mib":1172.55078125},{"t_s":104,"rss_mib":1172.69921875},{"t_s":105,"rss_mib":1172.69921875},{"t_s":107,"rss_mib":1172.70703125},{"t_s":108,"rss_mib":1172.70703125},{"t_s":110,"rss_mib":1172.78515625},{"t_s":111,"rss_mib":1172.78515625},{"t_s":112,"rss_mib":1172.796875},{"t_s":114,"rss_mib":1172.8046875},{"t_s":117,"rss_mib":1172.91015625},{"t_s":117,"rss_mib":1172.91015625},{"t_s":120,"rss_mib":1172.92578125},{"t_s":120,"rss_mib":1172.92578125},{"t_s":122,"rss_mib":1172.953125},{"t_s":123,"rss_mib":1172.953125},{"t_s":125,"rss_mib":1172.9765625},{"t_s":126,"rss_mib":1172.9765625},{"t_s":127,"rss_mib":1173.06640625},{"t_s":129,"rss_mib":1173.06640625},{"t_s":131,"rss_mib":1173.0703125},{"t_s":132,"rss_mib":1173.0703125},{"t_s":135,"rss_mib":1173.1953125},{"t_s":135,"rss_mib":1173.1953125},{"t_s":137,"rss_mib":1173.2265625},{"t_s":138,"rss_mib":1173.2265625},{"t_s":140,"rss_mib":1173.24609375},{"t_s":141,"rss_mib":1173.24609375},{"t_s":142,"rss_mib":1173.25},{"t_s":144,"rss_mib":1173.25},{"t_s":147,"rss_mib":1173.25},{"t_s":150,"rss_mib":1173.25},{"t_s":152,"rss_mib":1173.2578125},{"t_s":153,"rss_mib":1173.2578125},{"t_s":156,"rss_mib":1173.2578125},{"t_s":159,"rss_mib":1173.2578125},{"t_s":160,"rss_mib":1173.265625},{"t_s":162,"rss_mib":1173.265625},{"t_s":162,"rss_mib":1173.28125},{"t_s":165,"rss_mib":1173.2890625},{"t_s":165,"rss_mib":1173.29296875},{"t_s":168,"rss_mib":1173.3046875},{"t_s":169,"rss_mib":1170.81640625},{"t_s":171,"rss_mib":1170.81640625},{"t_s":173,"rss_mib":1170.828125},{"t_s":174,"rss_mib":1170.88671875},{"t_s":174,"rss_mib":1167.81640625},{"t_s":177,"rss_mib":1167.81640625},{"t_s":179,"rss_mib":1164.2109375},{"t_s":180,"rss_mib":1164.2109375},{"t_s":182,"rss_mib":1164.21484375},{"t_s":183,"rss_mib":1164.21484375},{"t_s":186,"rss_mib":1164.21484375},{"t_s":189,"rss_mib":1164.22265625},{"t_s":189,"rss_mib":1164.22265625},{"t_s":191,"rss_mib":1164.23828125},{"t_s":192,"rss_mib":1164.23828125},{"t_s":193,"rss_mib":1164.30078125},{"t_s":195,"rss_mib":1164.30078125},{"t_s":198,"rss_mib":1164.30078125},{"t_s":201,"rss_mib":1164.30078125},{"t_s":202,"rss_mib":1164.3046875},{"t_s":204,"rss_mib":1164.3046875},{"t_s":207,"rss_mib":1164.3046875},{"t_s":210,"rss_mib":1164.3046875},{"t_s":211,"rss_mib":1164.30859375},{"t_s":213,"rss_mib":1164.30859375},{"t_s":216,"rss_mib":1164.3125},{"t_s":219,"rss_mib":1164.3125},{"t_s":220,"rss_mib":1164.3203125},{"t_s":222,"rss_mib":1164.3203125},{"t_s":223,"rss_mib":1164.37890625},{"t_s":225,"rss_mib":1164.37890625},{"t_s":227,"rss_mib":1164.38671875},{"t_s":228,"rss_mib":1164.38671875},{"t_s":230,"rss_mib":1164.41796875},{"t_s":231,"rss_mib":1164.41796875},{"t_s":232,"rss_mib":1164.4296875},{"t_s":234,"rss_mib":1164.4296875},{"t_s":237,"rss_mib":1164.57421875},{"t_s":237,"rss_mib":1164.57421875},{"t_s":240,"rss_mib":1164.57421875},{"t_s":242,"rss_mib":1164.59765625},{"t_s":243,"rss_mib":1164.59765625},{"t_s":246,"rss_mib":1164.59765625},{"t_s":246,"rss_mib":1164.6015625},{"t_s":249,"rss_mib":1164.6015625},{"t_s":251,"rss_mib":1164.609375},{"t_s":252,"rss_mib":1164.609375},{"t_s":254,"rss_mib":1164.61328125},{"t_s":255,"rss_mib":1164.61328125},{"t_s":258,"rss_mib":1164.61328125},{"t_s":260,"rss_mib":1164.64453125},{"t_s":261,"rss_mib":1164.64453125},{"t_s":262,"rss_mib":1164.65625},{"t_s":264,"rss_mib":1164.65625},{"t_s":264,"rss_mib":1164.6796875},{"t_s":267,"rss_mib":1164.6796875},{"t_s":268,"rss_mib":1164.68359375},{"t_s":270,"rss_mib":1164.68359375},{"t_s":272,"rss_mib":1164.69140625},{"t_s":273,"rss_mib":1164.69140625},{"t_s":276,"rss_mib":1164.703125},{"t_s":276,"rss_mib":1164.703125},{"t_s":279,"rss_mib":1164.703125},{"t_s":282,"rss_mib":1164.703125},{"t_s":284,"rss_mib":1165.4140625},{"t_s":285,"rss_mib":1165.4140625},{"t_s":285,"rss_mib":1165.4375},{"t_s":288,"rss_mib":1165.4375},{"t_s":289,"rss_mib":1165.5},{"t_s":291,"rss_mib":1165.5},{"t_s":293,"rss_mib":1165.5234375},{"t_s":294,"rss_mib":1165.59765625},{"t_s":294,"rss_mib":1165.6015625},{"t_s":297,"rss_mib":1165.6015625},{"t_s":299,"rss_mib":1165.61328125},{"t_s":300,"rss_mib":1165.61328125},{"t_s":302,"rss_mib":1165.6171875},{"t_s":303,"rss_mib":1165.6171875},{"t_s":306,"rss_mib":1165.6171875},{"t_s":309,"rss_mib":1165.62109375},{"t_s":309,"rss_mib":1165.62109375},{"t_s":312,"rss_mib":1165.62109375},{"t_s":315,"rss_mib":1165.62109375},{"t_s":318,"rss_mib":1165.6484375},{"t_s":319,"rss_mib":1165.6875},{"t_s":321,"rss_mib":1165.6875},{"t_s":324,"rss_mib":1165.69140625},{"t_s":324,"rss_mib":1165.69140625},{"t_s":327,"rss_mib":1165.69140625},{"t_s":330,"rss_mib":1165.69140625},{"t_s":333,"rss_mib":1165.69140625},{"t_s":336,"rss_mib":1165.69140625},{"t_s":339,"rss_mib":1165.69140625},{"t_s":342,"rss_mib":1165.6953125},{"t_s":342,"rss_mib":1165.6953125},{"t_s":342,"rss_mib":1165.828125},{"t_s":345,"rss_mib":1165.828125},{"t_s":347,"rss_mib":1165.83203125},{"t_s":348,"rss_mib":1165.83203125},{"t_s":349,"rss_mib":1165.84765625},{"t_s":351,"rss_mib":1165.84765625},{"t_s":353,"rss_mib":1165.86328125},{"t_s":354,"rss_mib":1165.921875},{"t_s":357,"rss_mib":1165.921875},{"t_s":360,"rss_mib":1165.921875}],"idle_rss_series":[{"t_s":0,"rss_mib":1150.8359375},{"t_s":0,"rss_mib":1150.8984375},{"t_s":0,"rss_mib":1150.90625},{"t_s":0,"rss_mib":1150.90625},{"t_s":1,"rss_mib":1150.90625},{"t_s":1,"rss_mib":1150.9140625},{"t_s":1,"rss_mib":1150.9140625},{"t_s":2,"rss_mib":1150.921875},{"t_s":2,"rss_mib":1150.921875},{"t_s":2,"rss_mib":1151.03515625},{"t_s":2,"rss_mib":1151.03515625},{"t_s":3,"rss_mib":1151.03515625},{"t_s":3,"rss_mib":1151.03515625},{"t_s":4,"rss_mib":1151.0546875},{"t_s":4,"rss_mib":1151.0546875},{"t_s":4,"rss_mib":1151.24609375},{"t_s":4,"rss_mib":1151.24609375},{"t_s":5,"rss_mib":1151.24609375},{"t_s":5,"rss_mib":1151.25390625},{"t_s":5,"rss_mib":1151.25390625},{"t_s":6,"rss_mib":1151.26171875},{"t_s":6,"rss_mib":1151.26171875},{"t_s":6,"rss_mib":1151.26171875},{"t_s":7,"rss_mib":1151.26171875},{"t_s":7,"rss_mib":1151.27734375},{"t_s":7,"rss_mib":1151.27734375},{"t_s":8,"rss_mib":1151.27734375},{"t_s":8,"rss_mib":1151.27734375},{"t_s":9,"rss_mib":1151.28125},{"t_s":9,"rss_mib":1151.28125},{"t_s":9,"rss_mib":1151.2890625},{"t_s":9,"rss_mib":1151.2890625},{"t_s":10,"rss_mib":1151.2890625},{"t_s":10,"rss_mib":1151.296875},{"t_s":10,"rss_mib":1151.296875},{"t_s":11,"rss_mib":1151.296875},{"t_s":11,"rss_mib":1151.3046875},{"t_s":11,"rss_mib":1151.3046875},{"t_s":12,"rss_mib":1151.3046875},{"t_s":12,"rss_mib":1151.31640625},{"t_s":12,"rss_mib":1151.31640625},{"t_s":13,"rss_mib":1151.32421875},{"t_s":13,"rss_mib":1151.33203125},{"t_s":13,"rss_mib":1151.33984375},{"t_s":13,"rss_mib":1151.33984375},{"t_s":14,"rss_mib":1151.33984375},{"t_s":14,"rss_mib":1151.33984375},{"t_s":15,"rss_mib":1151.34765625},{"t_s":15,"rss_mib":1151.34765625},{"t_s":15,"rss_mib":1151.34765625},{"t_s":16,"rss_mib":1151.359375},{"t_s":16,"rss_mib":1151.359375},{"t_s":16,"rss_mib":1151.3671875},{"t_s":16,"rss_mib":1151.3671875},{"t_s":17,"rss_mib":1151.375},{"t_s":17,"rss_mib":1151.3828125},{"t_s":17,"rss_mib":1151.3828125},{"t_s":18,"rss_mib":1151.3828125},{"t_s":18,"rss_mib":1151.390625},{"t_s":18,"rss_mib":1151.390625},{"t_s":19,"rss_mib":1151.3984375},{"t_s":19,"rss_mib":1151.3984375},{"t_s":19,"rss_mib":1151.3984375},{"t_s":20,"rss_mib":1151.41015625},{"t_s":20,"rss_mib":1151.41015625},{"t_s":20,"rss_mib":1151.41796875},{"t_s":20,"rss_mib":1151.41796875},{"t_s":21,"rss_mib":1151.41796875},{"t_s":21,"rss_mib":1151.42578125},{"t_s":22,"rss_mib":1151.42578125},{"t_s":22,"rss_mib":1151.42578125},{"t_s":22,"rss_mib":1151.4296875},{"t_s":22,"rss_mib":1151.4296875},{"t_s":23,"rss_mib":1151.4453125},{"t_s":23,"rss_mib":1151.4453125},{"t_s":23,"rss_mib":1151.4453125},{"t_s":24,"rss_mib":1151.4453125},{"t_s":24,"rss_mib":1151.45703125},{"t_s":24,"rss_mib":1151.46484375},{"t_s":25,"rss_mib":1151.484375},{"t_s":25,"rss_mib":1151.484375},{"t_s":25,"rss_mib":1151.484375},{"t_s":26,"rss_mib":1151.484375},{"t_s":26,"rss_mib":1151.5},{"t_s":26,"rss_mib":1151.5},{"t_s":27,"rss_mib":1151.5078125},{"t_s":27,"rss_mib":1151.515625},{"t_s":27,"rss_mib":1151.515625},{"t_s":28,"rss_mib":1151.52734375},{"t_s":28,"rss_mib":1151.52734375},{"t_s":28,"rss_mib":1151.52734375},{"t_s":29,"rss_mib":1151.53125},{"t_s":29,"rss_mib":1151.53125},{"t_s":29,"rss_mib":1151.53125},{"t_s":30,"rss_mib":1151.5390625},{"t_s":30,"rss_mib":1151.546875},{"t_s":30,"rss_mib":1151.546875},{"t_s":31,"rss_mib":1151.5546875},{"t_s":31,"rss_mib":1151.5546875},{"t_s":31,"rss_mib":1151.5546875},{"t_s":32,"rss_mib":1151.5546875},{"t_s":32,"rss_mib":1151.5625},{"t_s":32,"rss_mib":1151.5625},{"t_s":33,"rss_mib":1151.5703125},{"t_s":33,"rss_mib":1151.578125},{"t_s":33,"rss_mib":1151.578125},{"t_s":34,"rss_mib":1151.578125},{"t_s":34,"rss_mib":1151.578125},{"t_s":35,"rss_mib":1151.5859375},{"t_s":35,"rss_mib":1151.5859375},{"t_s":35,"rss_mib":1151.59375},{"t_s":35,"rss_mib":1151.59375},{"t_s":36,"rss_mib":1151.6015625},{"t_s":36,"rss_mib":1151.6015625},{"t_s":36,"rss_mib":1151.609375},{"t_s":36,"rss_mib":1151.609375},{"t_s":37,"rss_mib":1151.6171875},{"t_s":37,"rss_mib":1151.6171875},{"t_s":37,"rss_mib":1151.6171875},{"t_s":38,"rss_mib":1151.6171875},{"t_s":38,"rss_mib":1151.6171875},{"t_s":39,"rss_mib":1151.62890625},{"t_s":39,"rss_mib":1151.63671875},{"t_s":39,"rss_mib":1151.64453125},{"t_s":40,"rss_mib":1151.64453125},{"t_s":41,"rss_mib":1151.64453125},{"t_s":41,"rss_mib":1151.65234375},{"t_s":41,"rss_mib":1151.65234375},{"t_s":41,"rss_mib":1151.65234375},{"t_s":42,"rss_mib":1151.66015625},{"t_s":42,"rss_mib":1151.66015625},{"t_s":42,"rss_mib":1151.94140625},{"t_s":42,"rss_mib":1151.94140625},{"t_s":43,"rss_mib":1151.94140625},{"t_s":44,"rss_mib":1151.94140625},{"t_s":44,"rss_mib":1151.94921875},{"t_s":44,"rss_mib":1151.94921875},{"t_s":44,"rss_mib":1151.94921875},{"t_s":45,"rss_mib":1151.9609375},{"t_s":45,"rss_mib":1151.96875},{"t_s":45,"rss_mib":1151.96875},{"t_s":46,"rss_mib":1151.9765625},{"t_s":46,"rss_mib":1151.9765625},{"t_s":47,"rss_mib":1151.9765625},{"t_s":47,"rss_mib":1151.9921875},{"t_s":47,"rss_mib":1151.9921875},{"t_s":47,"rss_mib":1151.9921875},{"t_s":48,"rss_mib":1152.01171875},{"t_s":48,"rss_mib":1152.01171875},{"t_s":48,"rss_mib":1152.01953125},{"t_s":48,"rss_mib":1152.01953125},{"t_s":49,"rss_mib":1152.0234375},{"t_s":49,"rss_mib":1152.0234375},{"t_s":49,"rss_mib":1152.03125},{"t_s":50,"rss_mib":1152.03125},{"t_s":50,"rss_mib":1152.0390625},{"t_s":50,"rss_mib":1152.0390625},{"t_s":50,"rss_mib":1152.0390625},{"t_s":51,"rss_mib":1152.05078125},{"t_s":51,"rss_mib":1152.05078125},{"t_s":51,"rss_mib":1152.05078125},{"t_s":52,"rss_mib":1152.0625},{"t_s":52,"rss_mib":1152.0625},{"t_s":52,"rss_mib":1152.07421875},{"t_s":52,"rss_mib":1152.07421875},{"t_s":53,"rss_mib":1152.109375},{"t_s":53,"rss_mib":1152.109375},{"t_s":53,"rss_mib":1152.109375},{"t_s":54,"rss_mib":1152.21484375},{"t_s":54,"rss_mib":1152.21484375},{"t_s":54,"rss_mib":1152.22265625},{"t_s":54,"rss_mib":1152.22265625},{"t_s":55,"rss_mib":1152.22265625},{"t_s":55,"rss_mib":1152.22265625},{"t_s":56,"rss_mib":1152.23046875},{"t_s":56,"rss_mib":1152.23046875},{"t_s":56,"rss_mib":1152.23828125},{"t_s":56,"rss_mib":1152.23828125},{"t_s":57,"rss_mib":1152.25},{"t_s":57,"rss_mib":1152.25},{"t_s":57,"rss_mib":1152.25},{"t_s":58,"rss_mib":1152.2578125},{"t_s":58,"rss_mib":1152.2578125},{"t_s":58,"rss_mib":1152.2578125},{"t_s":59,"rss_mib":1152.265625},{"t_s":59,"rss_mib":1152.26953125},{"t_s":59,"rss_mib":1152.27734375},{"t_s":59,"rss_mib":1152.27734375}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.010192582,"cost":6.034643158,"memory":433.898082328,"streaming":0.000002851,"streams_sustained":0.00000203,"throughput":339.551153024},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"perf.frontier.1ms.concurrency":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.p99_us":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.rps":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.samples":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"stream.added_gap_p50_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_gap_p99_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p50_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p99_us":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_direct_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_gw_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_fps":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_headroom":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_mock_ceiling":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_direct_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_gw_samples":{"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here"}}}}},"bedrock":{"configurable":true,"served":true,"egress_config":null,"serve_error":"","cells":{"openai-responses":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1/responses","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"gemini":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1beta/models/anthropic.claude-3-sonnet-20240229-v1:0:generateContent","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"cohere":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v2/chat","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"bedrock":{"served":true,"reason":null,"status":"","path":"/model/anthropic.claude-3-sonnet-20240229-v1:0/converse","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":727,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":861,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":889,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":778.1517645882627,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1285.0963597430407,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.9697792765122447,"certified":true,"suppressed":false},"cost_window_ok":{"value":30007,"certified":true,"suppressed":false},"cost_window_rps":{"value":5000,"certified":true,"suppressed":false},"cost_threads":{"value":143,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.005065484720231946,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":4996,"certified":true,"suppressed":false},"concurrency":8,"p99_us":3109,"first_disqualified_conc":16,"lower_bound":false,"samples":29985},{"bound_ms":10,"rps":{"value":5179,"certified":true,"suppressed":false},"concurrency":16,"p99_us":6789,"first_disqualified_conc":32,"lower_bound":false,"samples":31083},{"bound_ms":50,"rps":{"value":5403,"certified":true,"suppressed":false},"concurrency":64,"p99_us":21391,"first_disqualified_conc":256,"lower_bound":false,"samples":32432},{"bound_ms":100,"rps":{"value":5403,"certified":true,"suppressed":false},"concurrency":64,"p99_us":21391,"first_disqualified_conc":512,"lower_bound":false,"samples":32432},{"bound_ms":null,"rps":{"value":5403,"certified":true,"suppressed":false},"concurrency":64,"p99_us":21391,"first_disqualified_conc":null,"lower_bound":false,"samples":32432}],"sweep":[{"conc":1,"ok":7799,"rps":1299,"p99_us":1021,"fail":0},{"conc":1,"ok":7656,"rps":1275,"p99_us":1035,"fail":0},{"conc":1,"ok":7799,"rps":1299,"p99_us":1019,"fail":0},{"conc":2,"ok":15292,"rps":2548,"p99_us":1028,"fail":0},{"conc":2,"ok":15301,"rps":2549,"p99_us":1030,"fail":0},{"conc":2,"ok":8466,"rps":1410,"p99_us":1730,"fail":0},{"conc":4,"ok":23671,"rps":3944,"p99_us":1704,"fail":0},{"conc":4,"ok":16088,"rps":2680,"p99_us":2537,"fail":0},{"conc":4,"ok":27715,"rps":4618,"p99_us":1479,"fail":0},{"conc":8,"ok":29960,"rps":4992,"p99_us":3252,"fail":0},{"conc":8,"ok":29906,"rps":4983,"p99_us":3330,"fail":0},{"conc":8,"ok":29985,"rps":4996,"p99_us":3109,"fail":0},{"conc":16,"ok":30975,"rps":5161,"p99_us":6987,"fail":0},{"conc":16,"ok":31083,"rps":5179,"p99_us":6789,"fail":0},{"conc":16,"ok":30893,"rps":5148,"p99_us":6784,"fail":0},{"conc":32,"ok":31717,"rps":5285,"p99_us":15086,"fail":0},{"conc":32,"ok":31895,"rps":5314,"p99_us":12803,"fail":0},{"conc":32,"ok":31986,"rps":5330,"p99_us":10548,"fail":0},{"conc":64,"ok":32260,"rps":5376,"p99_us":22878,"fail":0},{"conc":64,"ok":32346,"rps":5389,"p99_us":21307,"fail":0},{"conc":64,"ok":32432,"rps":5403,"p99_us":21391,"fail":0},{"conc":128,"ok":32317,"rps":5385,"p99_us":37953,"fail":0},{"conc":128,"ok":32202,"rps":5366,"p99_us":40099,"fail":0},{"conc":128,"ok":31870,"rps":5310,"p99_us":38357,"fail":0},{"conc":256,"ok":30602,"rps":5099,"p99_us":69455,"fail":0},{"conc":256,"ok":30467,"rps":5077,"p99_us":70262,"fail":0},{"conc":256,"ok":30627,"rps":5103,"p99_us":71047,"fail":0},{"conc":512,"ok":29812,"rps":4967,"p99_us":129335,"fail":0},{"conc":512,"ok":29757,"rps":4958,"p99_us":134412,"fail":0},{"conc":512,"ok":29904,"rps":4982,"p99_us":142614,"fail":0},{"conc":1024,"ok":30108,"rps":5017,"p99_us":1371213,"fail":0},{"conc":1024,"ok":30486,"rps":5079,"p99_us":819800,"fail":0},{"conc":1024,"ok":30539,"rps":5089,"p99_us":418966,"fail":0},{"conc":2048,"ok":29973,"rps":4994,"p99_us":980751,"fail":0},{"conc":2048,"ok":30295,"rps":5047,"p99_us":934006,"fail":0},{"conc":2048,"ok":30077,"rps":5011,"p99_us":964646,"fail":0},{"conc":4096,"ok":30247,"rps":5039,"p99_us":882248,"fail":0},{"conc":4096,"ok":29956,"rps":4991,"p99_us":651180,"fail":0},{"conc":4096,"ok":30187,"rps":5030,"p99_us":917954,"fail":0},{"conc":8192,"ok":30190,"rps":5030,"p99_us":832761,"fail":0},{"conc":8192,"ok":30164,"rps":5026,"p99_us":859692,"fail":0},{"conc":8192,"ok":30168,"rps":5027,"p99_us":801073,"fail":0},{"conc":16384,"ok":30234,"rps":5038,"p99_us":661004,"fail":0},{"conc":16384,"ok":30255,"rps":5041,"p99_us":642617,"fail":0},{"conc":16384,"ok":30149,"rps":5024,"p99_us":788284,"fail":0},{"conc":32768,"ok":26907,"rps":4484,"p99_us":2950646,"fail":0},{"conc":32768,"ok":26538,"rps":4422,"p99_us":2596237,"fail":0},{"conc":32768,"ok":27366,"rps":4560,"p99_us":2608620,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1299,"certified":true,"suppressed":false},"p99_us":1021,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":2548,"certified":true,"suppressed":false},"p99_us":1030,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":3944,"certified":true,"suppressed":false},"p99_us":1704,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":4992,"certified":true,"suppressed":false},"p99_us":3252,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":5161,"certified":true,"suppressed":false},"p99_us":6789,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":5314,"certified":true,"suppressed":false},"p99_us":12803,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":5389,"certified":true,"suppressed":false},"p99_us":21391,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":5366,"certified":true,"suppressed":false},"p99_us":38357,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":5099,"certified":true,"suppressed":false},"p99_us":70262,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":4967,"certified":true,"suppressed":false},"p99_us":134412,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":5079,"certified":true,"suppressed":false},"p99_us":819800,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":5011,"certified":true,"suppressed":false},"p99_us":964646,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":5030,"certified":true,"suppressed":false},"p99_us":882248,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":5027,"certified":true,"suppressed":false},"p99_us":832761,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":5038,"certified":true,"suppressed":false},"p99_us":661004,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":4484,"certified":true,"suppressed":false},"p99_us":2608620,"windows":3,"clean_windows":3}],"reverify_note":"bedrock>bedrock is a same-dialect cell: there is no translation to prove, so a request-shape check upstream cannot distinguish a translating gateway from a forwarding one","c1_note":"the c=1 percentiles are taken over 7954 successful gateway round trip(s) and 263620 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":7954,"direct_c1_samples":263620},"stream":{"stream_served":"untestable","stream_error":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","reason":"untestable","stream_c1_note":null,"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"}},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1157.3984375,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1165.7421875,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1167.375,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1175.8359375,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1275.18359375,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.6896154469149554,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":0.6918641708843657,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1158.00390625},{"t_s":2,"rss_mib":1174.38671875},{"t_s":3,"rss_mib":1174.39453125},{"t_s":5,"rss_mib":1174.453125},{"t_s":6,"rss_mib":1174.45703125},{"t_s":8,"rss_mib":1174.59375},{"t_s":9,"rss_mib":1174.59375},{"t_s":11,"rss_mib":1174.61328125},{"t_s":12,"rss_mib":1174.61328125},{"t_s":14,"rss_mib":1174.7890625},{"t_s":15,"rss_mib":1174.7890625},{"t_s":16,"rss_mib":1174.80859375},{"t_s":18,"rss_mib":1174.80859375},{"t_s":19,"rss_mib":1175.02734375},{"t_s":21,"rss_mib":1175.02734375},{"t_s":23,"rss_mib":1175.08984375},{"t_s":24,"rss_mib":1175.08984375},{"t_s":26,"rss_mib":1175.09765625},{"t_s":27,"rss_mib":1175.09765625},{"t_s":29,"rss_mib":1175.109375},{"t_s":30,"rss_mib":1175.109375},{"t_s":30,"rss_mib":1175.15234375},{"t_s":33,"rss_mib":1175.15234375},{"t_s":35,"rss_mib":1175.18359375},{"t_s":36,"rss_mib":1175.18359375},{"t_s":38,"rss_mib":1175.1953125},{"t_s":39,"rss_mib":1175.1953125},{"t_s":41,"rss_mib":1175.4296875},{"t_s":42,"rss_mib":1175.43359375},{"t_s":44,"rss_mib":1175.51171875},{"t_s":45,"rss_mib":1175.51171875},{"t_s":46,"rss_mib":1175.51953125},{"t_s":48,"rss_mib":1175.51953125},{"t_s":50,"rss_mib":1175.55859375},{"t_s":51,"rss_mib":1175.55859375},{"t_s":53,"rss_mib":1175.625},{"t_s":54,"rss_mib":1175.8359375},{"t_s":54,"rss_mib":1171.26953125},{"t_s":57,"rss_mib":1171.26953125},{"t_s":59,"rss_mib":1171.38671875},{"t_s":60,"rss_mib":1171.38671875},{"t_s":60,"rss_mib":1171.4140625},{"t_s":63,"rss_mib":1171.4140625},{"t_s":66,"rss_mib":1171.4140625},{"t_s":68,"rss_mib":1171.421875},{"t_s":69,"rss_mib":1171.421875},{"t_s":69,"rss_mib":1171.42578125},{"t_s":72,"rss_mib":1171.42578125},{"t_s":74,"rss_mib":1171.4296875},{"t_s":75,"rss_mib":1171.4296875},{"t_s":78,"rss_mib":1171.4296875},{"t_s":79,"rss_mib":1171.43359375},{"t_s":81,"rss_mib":1171.4375},{"t_s":84,"rss_mib":1171.46875},{"t_s":84,"rss_mib":1171.46875},{"t_s":87,"rss_mib":1171.46875},{"t_s":89,"rss_mib":1171.47265625},{"t_s":90,"rss_mib":1171.47265625},{"t_s":91,"rss_mib":1171.48828125},{"t_s":93,"rss_mib":1171.48828125},{"t_s":95,"rss_mib":1171.50390625},{"t_s":96,"rss_mib":1171.53515625},{"t_s":98,"rss_mib":1171.58984375},{"t_s":99,"rss_mib":1171.58984375},{"t_s":101,"rss_mib":1171.79296875},{"t_s":102,"rss_mib":1171.79296875},{"t_s":104,"rss_mib":1171.80859375},{"t_s":105,"rss_mib":1171.80859375},{"t_s":107,"rss_mib":1171.81640625},{"t_s":108,"rss_mib":1171.81640625},{"t_s":109,"rss_mib":1171.82421875},{"t_s":111,"rss_mib":1171.83203125},{"t_s":114,"rss_mib":1171.9296875},{"t_s":114,"rss_mib":1171.9296875},{"t_s":116,"rss_mib":1171.95703125},{"t_s":117,"rss_mib":1171.95703125},{"t_s":119,"rss_mib":1171.97265625},{"t_s":120,"rss_mib":1171.97265625},{"t_s":122,"rss_mib":1171.9921875},{"t_s":123,"rss_mib":1171.99609375},{"t_s":125,"rss_mib":1172.00390625},{"t_s":126,"rss_mib":1172.00390625},{"t_s":128,"rss_mib":1172.03125},{"t_s":129,"rss_mib":1172.03125},{"t_s":131,"rss_mib":1172.0703125},{"t_s":132,"rss_mib":1172.07421875},{"t_s":133,"rss_mib":1172.109375},{"t_s":135,"rss_mib":1172.109375},{"t_s":137,"rss_mib":1172.12109375},{"t_s":138,"rss_mib":1172.12109375},{"t_s":140,"rss_mib":1172.19921875},{"t_s":141,"rss_mib":1172.19921875},{"t_s":143,"rss_mib":1172.2109375},{"t_s":144,"rss_mib":1172.234375},{"t_s":147,"rss_mib":1172.234375},{"t_s":149,"rss_mib":1172.23828125},{"t_s":150,"rss_mib":1172.23828125},{"t_s":150,"rss_mib":1172.24609375},{"t_s":153,"rss_mib":1172.24609375},{"t_s":153,"rss_mib":1172.25390625},{"t_s":156,"rss_mib":1172.25390625},{"t_s":159,"rss_mib":1172.25390625},{"t_s":160,"rss_mib":1172.40234375},{"t_s":162,"rss_mib":1172.40234375},{"t_s":165,"rss_mib":1172.40234375},{"t_s":165,"rss_mib":1172.41015625},{"t_s":170,"rss_mib":1172.42578125},{"t_s":171,"rss_mib":1170.64453125},{"t_s":171,"rss_mib":1170.72265625},{"t_s":173,"rss_mib":1170.7265625},{"t_s":175,"rss_mib":1170.7890625},{"t_s":175,"rss_mib":1167.01171875},{"t_s":178,"rss_mib":1167.015625},{"t_s":179,"rss_mib":1165.18359375},{"t_s":180,"rss_mib":1165.18359375},{"t_s":182,"rss_mib":1165.1875},{"t_s":183,"rss_mib":1165.1875},{"t_s":185,"rss_mib":1165.19921875},{"t_s":186,"rss_mib":1165.203125},{"t_s":188,"rss_mib":1165.2109375},{"t_s":189,"rss_mib":1165.2109375},{"t_s":192,"rss_mib":1165.2109375},{"t_s":195,"rss_mib":1165.2109375},{"t_s":198,"rss_mib":1165.234375},{"t_s":199,"rss_mib":1165.25},{"t_s":201,"rss_mib":1165.25},{"t_s":203,"rss_mib":1165.2578125},{"t_s":204,"rss_mib":1165.2578125},{"t_s":207,"rss_mib":1165.2578125},{"t_s":209,"rss_mib":1165.2734375},{"t_s":210,"rss_mib":1165.2734375},{"t_s":213,"rss_mib":1165.2734375},{"t_s":213,"rss_mib":1165.28125},{"t_s":216,"rss_mib":1165.28125},{"t_s":218,"rss_mib":1165.29296875},{"t_s":219,"rss_mib":1165.29296875},{"t_s":220,"rss_mib":1165.375},{"t_s":222,"rss_mib":1165.375},{"t_s":223,"rss_mib":1165.3828125},{"t_s":225,"rss_mib":1165.3828125},{"t_s":226,"rss_mib":1165.390625},{"t_s":228,"rss_mib":1165.390625},{"t_s":231,"rss_mib":1165.48828125},{"t_s":231,"rss_mib":1165.4921875},{"t_s":233,"rss_mib":1165.5078125},{"t_s":234,"rss_mib":1165.5234375},{"t_s":236,"rss_mib":1165.62109375},{"t_s":237,"rss_mib":1165.62109375},{"t_s":238,"rss_mib":1165.62890625},{"t_s":240,"rss_mib":1165.62890625},{"t_s":242,"rss_mib":1165.640625},{"t_s":243,"rss_mib":1165.640625},{"t_s":245,"rss_mib":1165.6484375},{"t_s":246,"rss_mib":1165.6484375},{"t_s":248,"rss_mib":1165.7109375},{"t_s":249,"rss_mib":1165.71875},{"t_s":249,"rss_mib":1165.7265625},{"t_s":252,"rss_mib":1165.7265625},{"t_s":255,"rss_mib":1165.75},{"t_s":255,"rss_mib":1165.75},{"t_s":257,"rss_mib":1165.7578125},{"t_s":258,"rss_mib":1165.7578125},{"t_s":261,"rss_mib":1165.78515625},{"t_s":261,"rss_mib":1165.78515625},{"t_s":263,"rss_mib":1165.79296875},{"t_s":264,"rss_mib":1165.79296875},{"t_s":267,"rss_mib":1165.82421875},{"t_s":267,"rss_mib":1165.82421875},{"t_s":270,"rss_mib":1165.83203125},{"t_s":270,"rss_mib":1165.83203125},{"t_s":272,"rss_mib":1165.83984375},{"t_s":273,"rss_mib":1165.83984375},{"t_s":275,"rss_mib":1165.85546875},{"t_s":276,"rss_mib":1165.85546875},{"t_s":277,"rss_mib":1165.89453125},{"t_s":279,"rss_mib":1165.89453125},{"t_s":280,"rss_mib":1166.046875},{"t_s":282,"rss_mib":1166.046875},{"t_s":285,"rss_mib":1166.046875},{"t_s":286,"rss_mib":1166.05859375},{"t_s":288,"rss_mib":1166.0703125},{"t_s":291,"rss_mib":1166.21875},{"t_s":291,"rss_mib":1166.21875},{"t_s":293,"rss_mib":1166.2265625},{"t_s":294,"rss_mib":1166.25},{"t_s":295,"rss_mib":1166.328125},{"t_s":297,"rss_mib":1166.328125},{"t_s":300,"rss_mib":1166.3515625},{"t_s":300,"rss_mib":1166.3515625},{"t_s":303,"rss_mib":1166.3515625},{"t_s":306,"rss_mib":1166.3515625},{"t_s":309,"rss_mib":1166.3515625},{"t_s":312,"rss_mib":1166.3515625},{"t_s":315,"rss_mib":1166.375},{"t_s":316,"rss_mib":1166.4140625},{"t_s":318,"rss_mib":1166.4140625},{"t_s":319,"rss_mib":1166.41796875},{"t_s":321,"rss_mib":1166.41796875},{"t_s":323,"rss_mib":1166.42578125},{"t_s":324,"rss_mib":1166.42578125},{"t_s":327,"rss_mib":1166.4296875},{"t_s":327,"rss_mib":1166.4296875},{"t_s":329,"rss_mib":1166.43359375},{"t_s":330,"rss_mib":1166.625},{"t_s":333,"rss_mib":1166.625},{"t_s":336,"rss_mib":1166.62890625},{"t_s":336,"rss_mib":1166.62890625},{"t_s":337,"rss_mib":1166.6328125},{"t_s":339,"rss_mib":1166.6328125},{"t_s":340,"rss_mib":1167.359375},{"t_s":342,"rss_mib":1167.359375},{"t_s":344,"rss_mib":1167.375},{"t_s":345,"rss_mib":1167.375},{"t_s":348,"rss_mib":1167.390625},{"t_s":348,"rss_mib":1167.390625},{"t_s":350,"rss_mib":1167.4140625},{"t_s":351,"rss_mib":1167.484375},{"t_s":352,"rss_mib":1167.4921875},{"t_s":354,"rss_mib":1167.4921875},{"t_s":357,"rss_mib":1167.50390625},{"t_s":357,"rss_mib":1167.50390625},{"t_s":359,"rss_mib":1167.52734375},{"t_s":360,"rss_mib":1167.52734375}],"idle_rss_series":[{"t_s":0,"rss_mib":1157.23828125},{"t_s":0,"rss_mib":1157.30078125},{"t_s":0,"rss_mib":1157.30078125},{"t_s":0,"rss_mib":1157.30078125},{"t_s":1,"rss_mib":1157.30078125},{"t_s":1,"rss_mib":1157.30078125},{"t_s":2,"rss_mib":1157.30078125},{"t_s":2,"rss_mib":1157.30078125},{"t_s":3,"rss_mib":1157.30078125},{"t_s":3,"rss_mib":1157.30078125},{"t_s":4,"rss_mib":1157.32421875},{"t_s":4,"rss_mib":1157.32421875},{"t_s":4,"rss_mib":1157.32421875},{"t_s":5,"rss_mib":1157.328125},{"t_s":5,"rss_mib":1157.328125},{"t_s":5,"rss_mib":1157.328125},{"t_s":6,"rss_mib":1157.328125},{"t_s":6,"rss_mib":1157.328125},{"t_s":7,"rss_mib":1157.328125},{"t_s":7,"rss_mib":1157.328125},{"t_s":8,"rss_mib":1157.328125},{"t_s":8,"rss_mib":1157.328125},{"t_s":9,"rss_mib":1157.328125},{"t_s":10,"rss_mib":1157.328125},{"t_s":10,"rss_mib":1157.328125},{"t_s":10,"rss_mib":1157.328125},{"t_s":11,"rss_mib":1157.328125},{"t_s":11,"rss_mib":1157.328125},{"t_s":12,"rss_mib":1157.328125},{"t_s":13,"rss_mib":1157.328125},{"t_s":13,"rss_mib":1157.328125},{"t_s":13,"rss_mib":1157.328125},{"t_s":14,"rss_mib":1157.33203125},{"t_s":14,"rss_mib":1157.33203125},{"t_s":14,"rss_mib":1157.33203125},{"t_s":15,"rss_mib":1157.33203125},{"t_s":15,"rss_mib":1157.33203125},{"t_s":16,"rss_mib":1157.3359375},{"t_s":16,"rss_mib":1157.3359375},{"t_s":16,"rss_mib":1157.3359375},{"t_s":17,"rss_mib":1157.3359375},{"t_s":17,"rss_mib":1157.3359375},{"t_s":18,"rss_mib":1157.3359375},{"t_s":18,"rss_mib":1157.3359375},{"t_s":19,"rss_mib":1157.33984375},{"t_s":19,"rss_mib":1157.33984375},{"t_s":20,"rss_mib":1157.33984375},{"t_s":20,"rss_mib":1157.33984375},{"t_s":20,"rss_mib":1157.33984375},{"t_s":21,"rss_mib":1157.33984375},{"t_s":21,"rss_mib":1157.33984375},{"t_s":22,"rss_mib":1157.33984375},{"t_s":23,"rss_mib":1157.33984375},{"t_s":23,"rss_mib":1157.33984375},{"t_s":23,"rss_mib":1157.33984375},{"t_s":24,"rss_mib":1157.3671875},{"t_s":24,"rss_mib":1157.3671875},{"t_s":24,"rss_mib":1157.3671875},{"t_s":25,"rss_mib":1157.375},{"t_s":25,"rss_mib":1157.375},{"t_s":26,"rss_mib":1157.375},{"t_s":26,"rss_mib":1157.375},{"t_s":26,"rss_mib":1157.375},{"t_s":27,"rss_mib":1157.375},{"t_s":27,"rss_mib":1157.375},{"t_s":28,"rss_mib":1157.3828125},{"t_s":28,"rss_mib":1157.3828125},{"t_s":29,"rss_mib":1157.3828125},{"t_s":29,"rss_mib":1157.390625},{"t_s":29,"rss_mib":1157.390625},{"t_s":29,"rss_mib":1157.3984375},{"t_s":30,"rss_mib":1157.3984375},{"t_s":30,"rss_mib":1157.3984375},{"t_s":30,"rss_mib":1157.40625},{"t_s":30,"rss_mib":1157.40625},{"t_s":31,"rss_mib":1157.4140625},{"t_s":31,"rss_mib":1157.4140625},{"t_s":31,"rss_mib":1157.4140625},{"t_s":32,"rss_mib":1157.421875},{"t_s":32,"rss_mib":1157.421875},{"t_s":33,"rss_mib":1157.4296875},{"t_s":33,"rss_mib":1157.4296875},{"t_s":33,"rss_mib":1157.4296875},{"t_s":34,"rss_mib":1157.4296875},{"t_s":34,"rss_mib":1157.4375},{"t_s":34,"rss_mib":1157.4375},{"t_s":35,"rss_mib":1157.4375},{"t_s":36,"rss_mib":1157.4375},{"t_s":36,"rss_mib":1157.4453125},{"t_s":36,"rss_mib":1157.4453125},{"t_s":36,"rss_mib":1157.4453125},{"t_s":37,"rss_mib":1157.4453125},{"t_s":37,"rss_mib":1157.453125},{"t_s":37,"rss_mib":1157.453125},{"t_s":38,"rss_mib":1157.453125},{"t_s":39,"rss_mib":1157.453125},{"t_s":39,"rss_mib":1157.45703125},{"t_s":39,"rss_mib":1157.45703125},{"t_s":39,"rss_mib":1157.4609375},{"t_s":40,"rss_mib":1157.4609375},{"t_s":40,"rss_mib":1157.46875},{"t_s":40,"rss_mib":1157.46875},{"t_s":40,"rss_mib":1157.69140625},{"t_s":41,"rss_mib":1157.69140625},{"t_s":42,"rss_mib":1157.69140625},{"t_s":42,"rss_mib":1157.69140625},{"t_s":43,"rss_mib":1157.69140625},{"t_s":43,"rss_mib":1157.69921875},{"t_s":43,"rss_mib":1157.69921875},{"t_s":43,"rss_mib":1157.69921875},{"t_s":44,"rss_mib":1157.69921875},{"t_s":45,"rss_mib":1157.69921875},{"t_s":45,"rss_mib":1157.69921875},{"t_s":45,"rss_mib":1157.71484375},{"t_s":46,"rss_mib":1157.72265625},{"t_s":46,"rss_mib":1157.72265625},{"t_s":46,"rss_mib":1157.72265625},{"t_s":47,"rss_mib":1157.72265625},{"t_s":47,"rss_mib":1157.72265625},{"t_s":48,"rss_mib":1157.73046875},{"t_s":48,"rss_mib":1157.73046875},{"t_s":48,"rss_mib":1157.73828125},{"t_s":49,"rss_mib":1157.73828125},{"t_s":49,"rss_mib":1157.73828125},{"t_s":49,"rss_mib":1157.7421875},{"t_s":49,"rss_mib":1157.7421875},{"t_s":50,"rss_mib":1157.76171875},{"t_s":50,"rss_mib":1157.76171875},{"t_s":50,"rss_mib":1157.76171875},{"t_s":51,"rss_mib":1157.87890625},{"t_s":51,"rss_mib":1157.87890625},{"t_s":51,"rss_mib":1157.88671875},{"t_s":52,"rss_mib":1157.88671875},{"t_s":52,"rss_mib":1157.89453125},{"t_s":52,"rss_mib":1157.89453125},{"t_s":53,"rss_mib":1157.89453125},{"t_s":53,"rss_mib":1157.89453125},{"t_s":53,"rss_mib":1157.90234375},{"t_s":53,"rss_mib":1157.90234375},{"t_s":54,"rss_mib":1157.91015625},{"t_s":54,"rss_mib":1157.91015625},{"t_s":55,"rss_mib":1157.91015625},{"t_s":55,"rss_mib":1157.91796875},{"t_s":55,"rss_mib":1157.91796875},{"t_s":56,"rss_mib":1157.92578125},{"t_s":56,"rss_mib":1157.93359375},{"t_s":56,"rss_mib":1157.93359375},{"t_s":56,"rss_mib":1157.93359375},{"t_s":57,"rss_mib":1157.94140625},{"t_s":57,"rss_mib":1157.94140625},{"t_s":58,"rss_mib":1157.94140625},{"t_s":58,"rss_mib":1157.9453125},{"t_s":58,"rss_mib":1157.9453125},{"t_s":58,"rss_mib":1157.953125},{"t_s":59,"rss_mib":1157.95703125},{"t_s":59,"rss_mib":1157.95703125},{"t_s":59,"rss_mib":1157.9609375},{"t_s":59,"rss_mib":1157.9609375}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.009518762,"cost":6.034281781,"memory":436.959955542,"streaming":0.000002503,"streams_sustained":0.000002158,"throughput":327.635908337},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"perf.frontier.1ms.concurrency":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.p99_us":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.rps":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.samples":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"stream.added_gap_p50_us":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_gap_p99_us":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p50_us":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p99_us":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_direct_samples":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_gw_samples":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_fps":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_headroom":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_mock_ceiling":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_direct_samples":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_gw_samples":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"}}},"openai":{"served":true,"reason":null,"status":"","path":"/v1/chat/completions","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":929,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":1235,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":1263,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":912.3461771232601,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1096.0751796572692,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.7513336009437701,"certified":true,"suppressed":false},"cost_window_ok":{"value":19828,"certified":true,"suppressed":false},"cost_window_rps":{"value":3304,"certified":true,"suppressed":false},"cost_threads":{"value":144,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.008069396812588258,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":4316,"certified":true,"suppressed":false},"concurrency":8,"p99_us":3304,"first_disqualified_conc":16,"lower_bound":false,"samples":25900},{"bound_ms":10,"rps":{"value":4440,"certified":true,"suppressed":false},"concurrency":16,"p99_us":5213,"first_disqualified_conc":32,"lower_bound":false,"samples":26646},{"bound_ms":50,"rps":{"value":4587,"certified":true,"suppressed":false},"concurrency":64,"p99_us":20787,"first_disqualified_conc":256,"lower_bound":false,"samples":27530},{"bound_ms":100,"rps":{"value":4587,"certified":true,"suppressed":false},"concurrency":64,"p99_us":20787,"first_disqualified_conc":512,"lower_bound":false,"samples":27530},{"bound_ms":null,"rps":{"value":4587,"certified":true,"suppressed":false},"concurrency":64,"p99_us":20787,"first_disqualified_conc":null,"lower_bound":false,"samples":27530}],"sweep":[{"conc":1,"ok":6289,"rps":1047,"p99_us":1238,"fail":0},{"conc":1,"ok":6230,"rps":1038,"p99_us":1258,"fail":0},{"conc":1,"ok":6279,"rps":1046,"p99_us":1266,"fail":0},{"conc":2,"ok":11525,"rps":1920,"p99_us":1314,"fail":0},{"conc":2,"ok":6620,"rps":1103,"p99_us":2200,"fail":0},{"conc":2,"ok":11975,"rps":1995,"p99_us":1281,"fail":0},{"conc":4,"ok":13235,"rps":2205,"p99_us":2207,"fail":0},{"conc":4,"ok":18946,"rps":3157,"p99_us":2139,"fail":0},{"conc":4,"ok":12806,"rps":2133,"p99_us":3449,"fail":0},{"conc":8,"ok":25900,"rps":4316,"p99_us":3304,"fail":0},{"conc":8,"ok":25383,"rps":4229,"p99_us":3407,"fail":0},{"conc":8,"ok":19913,"rps":3318,"p99_us":3394,"fail":0},{"conc":16,"ok":26646,"rps":4440,"p99_us":5213,"fail":0},{"conc":16,"ok":26629,"rps":4437,"p99_us":6490,"fail":0},{"conc":16,"ok":26503,"rps":4416,"p99_us":5329,"fail":0},{"conc":32,"ok":27275,"rps":4544,"p99_us":11855,"fail":0},{"conc":32,"ok":27294,"rps":4548,"p99_us":12949,"fail":0},{"conc":32,"ok":27305,"rps":4549,"p99_us":10048,"fail":0},{"conc":64,"ok":27468,"rps":4577,"p99_us":23571,"fail":0},{"conc":64,"ok":27530,"rps":4587,"p99_us":20787,"fail":0},{"conc":64,"ok":27520,"rps":4586,"p99_us":20631,"fail":0},{"conc":128,"ok":27397,"rps":4565,"p99_us":37933,"fail":0},{"conc":128,"ok":27435,"rps":4571,"p99_us":39988,"fail":0},{"conc":128,"ok":27421,"rps":4569,"p99_us":36460,"fail":0},{"conc":256,"ok":26334,"rps":4387,"p99_us":79493,"fail":0},{"conc":256,"ok":26294,"rps":4381,"p99_us":79784,"fail":0},{"conc":256,"ok":26327,"rps":4386,"p99_us":75207,"fail":0},{"conc":512,"ok":25695,"rps":4281,"p99_us":152220,"fail":0},{"conc":512,"ok":25739,"rps":4289,"p99_us":157463,"fail":0},{"conc":512,"ok":25768,"rps":4293,"p99_us":162923,"fail":0},{"conc":1024,"ok":26044,"rps":4339,"p99_us":1006016,"fail":0},{"conc":1024,"ok":25911,"rps":4317,"p99_us":1001265,"fail":0},{"conc":1024,"ok":25904,"rps":4316,"p99_us":299140,"fail":0},{"conc":2048,"ok":25198,"rps":4198,"p99_us":1094242,"fail":0},{"conc":2048,"ok":25398,"rps":4232,"p99_us":1019431,"fail":0},{"conc":2048,"ok":25156,"rps":4191,"p99_us":1074756,"fail":0},{"conc":4096,"ok":25477,"rps":4245,"p99_us":1006169,"fail":0},{"conc":4096,"ok":25810,"rps":4301,"p99_us":1950130,"fail":0},{"conc":4096,"ok":25415,"rps":4234,"p99_us":971574,"fail":0},{"conc":8192,"ok":25910,"rps":4317,"p99_us":2341125,"fail":0},{"conc":8192,"ok":25335,"rps":4222,"p99_us":747627,"fail":0},{"conc":8192,"ok":25585,"rps":4263,"p99_us":1081959,"fail":0},{"conc":16384,"ok":25405,"rps":4233,"p99_us":779185,"fail":0},{"conc":16384,"ok":25432,"rps":4237,"p99_us":1012580,"fail":0},{"conc":16384,"ok":25517,"rps":4251,"p99_us":669767,"fail":0},{"conc":32768,"ok":21705,"rps":3616,"p99_us":3007030,"fail":0},{"conc":32768,"ok":21760,"rps":3625,"p99_us":2651335,"fail":0},{"conc":32768,"ok":23368,"rps":3893,"p99_us":2590372,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1046,"certified":true,"suppressed":false},"p99_us":1258,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":1920,"certified":true,"suppressed":false},"p99_us":1314,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":2205,"certified":true,"suppressed":false},"p99_us":2207,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":4229,"certified":true,"suppressed":false},"p99_us":3394,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":4437,"certified":true,"suppressed":false},"p99_us":5329,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":4548,"certified":true,"suppressed":false},"p99_us":11855,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":4586,"certified":true,"suppressed":false},"p99_us":20787,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":4569,"certified":true,"suppressed":false},"p99_us":37933,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":4386,"certified":true,"suppressed":false},"p99_us":79493,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":4289,"certified":true,"suppressed":false},"p99_us":157463,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":4317,"certified":true,"suppressed":false},"p99_us":1001265,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":4198,"certified":true,"suppressed":false},"p99_us":1074756,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":4245,"certified":true,"suppressed":false},"p99_us":1006169,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":4263,"certified":true,"suppressed":false},"p99_us":1081959,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":4237,"certified":true,"suppressed":false},"p99_us":779185,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":3625,"certified":true,"suppressed":false},"p99_us":2651335,"windows":3,"clean_windows":3}],"egress_reverified":true,"c1_note":"the c=1 percentiles are taken over 6162 successful gateway round trip(s) and 262294 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":6162,"direct_c1_samples":262294},"stream":{"stream_served":"untestable","stream_error":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","reason":"untestable","stream_c1_note":null,"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"}},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1158.6875,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1172.09765625,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1173.9765625,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1179.87109375,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1198.6875,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.9309596103361604,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":1.0752406083257144,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1159.4921875},{"t_s":2,"rss_mib":1178.3125},{"t_s":3,"rss_mib":1178.3125},{"t_s":5,"rss_mib":1178.35546875},{"t_s":6,"rss_mib":1178.359375},{"t_s":8,"rss_mib":1178.60546875},{"t_s":9,"rss_mib":1178.60546875},{"t_s":11,"rss_mib":1178.63671875},{"t_s":12,"rss_mib":1178.640625},{"t_s":14,"rss_mib":1178.85546875},{"t_s":15,"rss_mib":1178.86328125},{"t_s":16,"rss_mib":1178.91015625},{"t_s":18,"rss_mib":1178.91015625},{"t_s":20,"rss_mib":1178.94921875},{"t_s":21,"rss_mib":1178.94921875},{"t_s":23,"rss_mib":1178.97265625},{"t_s":24,"rss_mib":1178.9765625},{"t_s":25,"rss_mib":1179.10546875},{"t_s":27,"rss_mib":1179.10546875},{"t_s":29,"rss_mib":1179.14453125},{"t_s":30,"rss_mib":1179.14453125},{"t_s":32,"rss_mib":1179.20703125},{"t_s":33,"rss_mib":1179.20703125},{"t_s":34,"rss_mib":1179.2265625},{"t_s":36,"rss_mib":1179.2265625},{"t_s":37,"rss_mib":1179.2578125},{"t_s":39,"rss_mib":1179.26171875},{"t_s":41,"rss_mib":1179.30078125},{"t_s":42,"rss_mib":1179.30078125},{"t_s":44,"rss_mib":1179.48046875},{"t_s":45,"rss_mib":1179.48046875},{"t_s":47,"rss_mib":1179.51171875},{"t_s":48,"rss_mib":1179.51171875},{"t_s":49,"rss_mib":1179.765625},{"t_s":51,"rss_mib":1179.765625},{"t_s":53,"rss_mib":1179.77734375},{"t_s":54,"rss_mib":1179.77734375},{"t_s":55,"rss_mib":1179.87109375},{"t_s":57,"rss_mib":1179.87109375},{"t_s":59,"rss_mib":1177.91796875},{"t_s":60,"rss_mib":1177.91796875},{"t_s":62,"rss_mib":1177.9609375},{"t_s":63,"rss_mib":1177.9609375},{"t_s":63,"rss_mib":1178},{"t_s":66,"rss_mib":1178},{"t_s":68,"rss_mib":1178.0859375},{"t_s":69,"rss_mib":1178.0859375},{"t_s":69,"rss_mib":1178.09375},{"t_s":72,"rss_mib":1178.09375},{"t_s":72,"rss_mib":1178.09765625},{"t_s":75,"rss_mib":1178.09765625},{"t_s":76,"rss_mib":1178.1171875},{"t_s":78,"rss_mib":1178.1171875},{"t_s":81,"rss_mib":1178.1171875},{"t_s":83,"rss_mib":1178.12890625},{"t_s":84,"rss_mib":1178.12890625},{"t_s":86,"rss_mib":1178.18359375},{"t_s":87,"rss_mib":1178.18359375},{"t_s":88,"rss_mib":1178.19921875},{"t_s":90,"rss_mib":1178.19921875},{"t_s":92,"rss_mib":1178.22265625},{"t_s":93,"rss_mib":1178.22265625},{"t_s":94,"rss_mib":1178.23046875},{"t_s":96,"rss_mib":1178.23828125},{"t_s":97,"rss_mib":1178.25},{"t_s":99,"rss_mib":1178.25},{"t_s":101,"rss_mib":1178.265625},{"t_s":102,"rss_mib":1178.28515625},{"t_s":103,"rss_mib":1178.3515625},{"t_s":105,"rss_mib":1178.3515625},{"t_s":108,"rss_mib":1178.375},{"t_s":108,"rss_mib":1178.375},{"t_s":110,"rss_mib":1178.3828125},{"t_s":111,"rss_mib":1178.3828125},{"t_s":113,"rss_mib":1178.41015625},{"t_s":114,"rss_mib":1178.41015625},{"t_s":115,"rss_mib":1178.5390625},{"t_s":117,"rss_mib":1178.5390625},{"t_s":119,"rss_mib":1178.5703125},{"t_s":120,"rss_mib":1178.58984375},{"t_s":122,"rss_mib":1178.6015625},{"t_s":123,"rss_mib":1178.609375},{"t_s":126,"rss_mib":1178.625},{"t_s":126,"rss_mib":1178.625},{"t_s":129,"rss_mib":1178.63671875},{"t_s":129,"rss_mib":1178.63671875},{"t_s":132,"rss_mib":1178.65234375},{"t_s":132,"rss_mib":1178.65234375},{"t_s":135,"rss_mib":1178.6875},{"t_s":135,"rss_mib":1178.6875},{"t_s":138,"rss_mib":1178.73828125},{"t_s":138,"rss_mib":1178.73828125},{"t_s":140,"rss_mib":1178.765625},{"t_s":141,"rss_mib":1178.765625},{"t_s":143,"rss_mib":1178.78125},{"t_s":144,"rss_mib":1178.78125},{"t_s":145,"rss_mib":1178.83203125},{"t_s":147,"rss_mib":1178.83203125},{"t_s":147,"rss_mib":1178.83984375},{"t_s":150,"rss_mib":1178.83984375},{"t_s":153,"rss_mib":1178.8515625},{"t_s":153,"rss_mib":1178.8515625},{"t_s":154,"rss_mib":1178.86328125},{"t_s":156,"rss_mib":1178.86328125},{"t_s":159,"rss_mib":1178.86328125},{"t_s":161,"rss_mib":1178.8671875},{"t_s":162,"rss_mib":1178.8671875},{"t_s":164,"rss_mib":1178.90234375},{"t_s":165,"rss_mib":1178.90234375},{"t_s":167,"rss_mib":1178.91015625},{"t_s":169,"rss_mib":1178.9453125},{"t_s":169,"rss_mib":1174.6796875},{"t_s":171,"rss_mib":1174.6796875},{"t_s":173,"rss_mib":1174.68359375},{"t_s":174,"rss_mib":1174.68359375},{"t_s":174,"rss_mib":1172.78515625},{"t_s":179,"rss_mib":1172.8359375},{"t_s":179,"rss_mib":1171.36328125},{"t_s":180,"rss_mib":1171.36328125},{"t_s":183,"rss_mib":1171.36328125},{"t_s":186,"rss_mib":1171.36328125},{"t_s":188,"rss_mib":1171.375},{"t_s":189,"rss_mib":1171.375},{"t_s":190,"rss_mib":1171.38671875},{"t_s":192,"rss_mib":1171.45703125},{"t_s":195,"rss_mib":1171.48828125},{"t_s":195,"rss_mib":1171.48828125},{"t_s":198,"rss_mib":1171.48828125},{"t_s":201,"rss_mib":1171.49609375},{"t_s":202,"rss_mib":1171.5},{"t_s":204,"rss_mib":1171.5},{"t_s":204,"rss_mib":1171.59375},{"t_s":207,"rss_mib":1171.59375},{"t_s":209,"rss_mib":1171.609375},{"t_s":210,"rss_mib":1171.609375},{"t_s":213,"rss_mib":1171.609375},{"t_s":216,"rss_mib":1171.609375},{"t_s":217,"rss_mib":1171.61328125},{"t_s":219,"rss_mib":1171.61328125},{"t_s":220,"rss_mib":1171.6171875},{"t_s":222,"rss_mib":1171.625},{"t_s":224,"rss_mib":1171.66796875},{"t_s":225,"rss_mib":1171.66796875},{"t_s":227,"rss_mib":1171.67578125},{"t_s":228,"rss_mib":1171.67578125},{"t_s":230,"rss_mib":1171.6796875},{"t_s":231,"rss_mib":1171.6796875},{"t_s":233,"rss_mib":1171.69921875},{"t_s":234,"rss_mib":1171.76171875},{"t_s":235,"rss_mib":1171.79296875},{"t_s":237,"rss_mib":1171.79296875},{"t_s":239,"rss_mib":1171.796875},{"t_s":240,"rss_mib":1171.796875},{"t_s":241,"rss_mib":1171.8046875},{"t_s":243,"rss_mib":1171.8046875},{"t_s":244,"rss_mib":1172.06640625},{"t_s":246,"rss_mib":1172.06640625},{"t_s":246,"rss_mib":1172.07421875},{"t_s":249,"rss_mib":1172.07421875},{"t_s":251,"rss_mib":1172.08203125},{"t_s":252,"rss_mib":1172.08203125},{"t_s":253,"rss_mib":1172.09765625},{"t_s":255,"rss_mib":1172.09765625},{"t_s":258,"rss_mib":1172.09765625},{"t_s":260,"rss_mib":1172.1015625},{"t_s":261,"rss_mib":1172.1015625},{"t_s":263,"rss_mib":1172.140625},{"t_s":264,"rss_mib":1172.140625},{"t_s":266,"rss_mib":1172.2578125},{"t_s":267,"rss_mib":1172.2578125},{"t_s":270,"rss_mib":1172.2890625},{"t_s":270,"rss_mib":1172.2890625},{"t_s":272,"rss_mib":1172.296875},{"t_s":273,"rss_mib":1172.296875},{"t_s":275,"rss_mib":1172.3125},{"t_s":276,"rss_mib":1172.36328125},{"t_s":279,"rss_mib":1172.37890625},{"t_s":279,"rss_mib":1172.37890625},{"t_s":281,"rss_mib":1172.3984375},{"t_s":282,"rss_mib":1172.3984375},{"t_s":284,"rss_mib":1172.5859375},{"t_s":285,"rss_mib":1172.5859375},{"t_s":287,"rss_mib":1172.6015625},{"t_s":288,"rss_mib":1172.6015625},{"t_s":289,"rss_mib":1172.609375},{"t_s":291,"rss_mib":1172.609375},{"t_s":294,"rss_mib":1172.76171875},{"t_s":294,"rss_mib":1172.76171875},{"t_s":296,"rss_mib":1172.87109375},{"t_s":298,"rss_mib":1172.875},{"t_s":299,"rss_mib":1172.75},{"t_s":300,"rss_mib":1172.75},{"t_s":302,"rss_mib":1172.76171875},{"t_s":303,"rss_mib":1172.76171875},{"t_s":305,"rss_mib":1172.86328125},{"t_s":306,"rss_mib":1172.86328125},{"t_s":307,"rss_mib":1172.9296875},{"t_s":309,"rss_mib":1172.9296875},{"t_s":312,"rss_mib":1172.9296875},{"t_s":315,"rss_mib":1172.9296875},{"t_s":316,"rss_mib":1172.93359375},{"t_s":318,"rss_mib":1172.9609375},{"t_s":319,"rss_mib":1172.9921875},{"t_s":321,"rss_mib":1172.9921875},{"t_s":324,"rss_mib":1172.9921875},{"t_s":325,"rss_mib":1173.01171875},{"t_s":327,"rss_mib":1173.01171875},{"t_s":329,"rss_mib":1173.01953125},{"t_s":330,"rss_mib":1173.01953125},{"t_s":332,"rss_mib":1173.02734375},{"t_s":333,"rss_mib":1173.02734375},{"t_s":336,"rss_mib":1173.02734375},{"t_s":337,"rss_mib":1173.16015625},{"t_s":339,"rss_mib":1173.16015625},{"t_s":341,"rss_mib":1173.1796875},{"t_s":342,"rss_mib":1173.1796875},{"t_s":344,"rss_mib":1173.9765625},{"t_s":345,"rss_mib":1173.9765625},{"t_s":348,"rss_mib":1173.9921875},{"t_s":348,"rss_mib":1173.9921875},{"t_s":351,"rss_mib":1173.9921875},{"t_s":354,"rss_mib":1174.078125},{"t_s":354,"rss_mib":1174.078125},{"t_s":355,"rss_mib":1174.08984375},{"t_s":357,"rss_mib":1174.08984375},{"t_s":359,"rss_mib":1174.3125},{"t_s":360,"rss_mib":1174.3125}],"idle_rss_series":[{"t_s":0,"rss_mib":1158.1875},{"t_s":0,"rss_mib":1158.25},{"t_s":0,"rss_mib":1158.25},{"t_s":0,"rss_mib":1158.25},{"t_s":1,"rss_mib":1158.25},{"t_s":1,"rss_mib":1158.25390625},{"t_s":1,"rss_mib":1158.26171875},{"t_s":2,"rss_mib":1158.2734375},{"t_s":2,"rss_mib":1158.2734375},{"t_s":2,"rss_mib":1158.37109375},{"t_s":2,"rss_mib":1158.375},{"t_s":3,"rss_mib":1158.375},{"t_s":3,"rss_mib":1158.375},{"t_s":4,"rss_mib":1158.3828125},{"t_s":4,"rss_mib":1158.3828125},{"t_s":4,"rss_mib":1158.38671875},{"t_s":4,"rss_mib":1158.38671875},{"t_s":5,"rss_mib":1158.38671875},{"t_s":5,"rss_mib":1158.39453125},{"t_s":5,"rss_mib":1158.3984375},{"t_s":6,"rss_mib":1158.3984375},{"t_s":6,"rss_mib":1158.3984375},{"t_s":7,"rss_mib":1158.3984375},{"t_s":7,"rss_mib":1158.4140625},{"t_s":7,"rss_mib":1158.4140625},{"t_s":8,"rss_mib":1158.4140625},{"t_s":8,"rss_mib":1158.421875},{"t_s":8,"rss_mib":1158.421875},{"t_s":9,"rss_mib":1158.4296875},{"t_s":10,"rss_mib":1158.4296875},{"t_s":10,"rss_mib":1158.4375},{"t_s":10,"rss_mib":1158.4453125},{"t_s":10,"rss_mib":1158.4453125},{"t_s":11,"rss_mib":1158.453125},{"t_s":11,"rss_mib":1158.453125},{"t_s":12,"rss_mib":1158.453125},{"t_s":13,"rss_mib":1158.453125},{"t_s":13,"rss_mib":1158.4609375},{"t_s":13,"rss_mib":1158.4609375},{"t_s":13,"rss_mib":1158.47265625},{"t_s":13,"rss_mib":1158.47265625},{"t_s":14,"rss_mib":1158.47265625},{"t_s":14,"rss_mib":1158.48046875},{"t_s":15,"rss_mib":1158.48046875},{"t_s":15,"rss_mib":1158.48828125},{"t_s":16,"rss_mib":1158.48828125},{"t_s":16,"rss_mib":1158.48828125},{"t_s":16,"rss_mib":1158.48828125},{"t_s":17,"rss_mib":1158.50390625},{"t_s":17,"rss_mib":1158.50390625},{"t_s":17,"rss_mib":1158.5234375},{"t_s":17,"rss_mib":1158.53125},{"t_s":18,"rss_mib":1158.53125},{"t_s":18,"rss_mib":1158.53125},{"t_s":19,"rss_mib":1158.53125},{"t_s":19,"rss_mib":1158.5390625},{"t_s":20,"rss_mib":1158.5390625},{"t_s":20,"rss_mib":1158.54296875},{"t_s":20,"rss_mib":1158.54296875},{"t_s":20,"rss_mib":1158.54296875},{"t_s":21,"rss_mib":1158.55078125},{"t_s":21,"rss_mib":1158.55859375},{"t_s":22,"rss_mib":1158.56640625},{"t_s":22,"rss_mib":1158.56640625},{"t_s":22,"rss_mib":1158.5703125},{"t_s":23,"rss_mib":1158.5703125},{"t_s":23,"rss_mib":1158.578125},{"t_s":23,"rss_mib":1158.578125},{"t_s":24,"rss_mib":1158.578125},{"t_s":24,"rss_mib":1158.6171875},{"t_s":24,"rss_mib":1158.6171875},{"t_s":25,"rss_mib":1158.640625},{"t_s":25,"rss_mib":1158.640625},{"t_s":26,"rss_mib":1158.640625},{"t_s":26,"rss_mib":1158.640625},{"t_s":26,"rss_mib":1158.6484375},{"t_s":27,"rss_mib":1158.65625},{"t_s":27,"rss_mib":1158.65625},{"t_s":27,"rss_mib":1158.65625},{"t_s":28,"rss_mib":1158.6640625},{"t_s":28,"rss_mib":1158.6640625},{"t_s":28,"rss_mib":1158.6796875},{"t_s":29,"rss_mib":1158.6796875},{"t_s":29,"rss_mib":1158.6796875},{"t_s":29,"rss_mib":1158.6875},{"t_s":30,"rss_mib":1158.6875},{"t_s":30,"rss_mib":1158.6875},{"t_s":30,"rss_mib":1158.6953125},{"t_s":30,"rss_mib":1158.703125},{"t_s":31,"rss_mib":1158.70703125},{"t_s":31,"rss_mib":1158.70703125},{"t_s":31,"rss_mib":1158.70703125},{"t_s":32,"rss_mib":1158.71484375},{"t_s":32,"rss_mib":1158.71484375},{"t_s":32,"rss_mib":1158.7265625},{"t_s":33,"rss_mib":1158.7265625},{"t_s":33,"rss_mib":1158.7265625},{"t_s":33,"rss_mib":1158.734375},{"t_s":33,"rss_mib":1158.734375},{"t_s":34,"rss_mib":1158.74609375},{"t_s":34,"rss_mib":1158.74609375},{"t_s":34,"rss_mib":1158.74609375},{"t_s":35,"rss_mib":1158.75390625},{"t_s":35,"rss_mib":1158.75390625},{"t_s":35,"rss_mib":1158.76171875},{"t_s":36,"rss_mib":1158.76171875},{"t_s":36,"rss_mib":1158.76171875},{"t_s":36,"rss_mib":1158.76171875},{"t_s":37,"rss_mib":1158.76953125},{"t_s":37,"rss_mib":1158.78515625},{"t_s":37,"rss_mib":1158.78515625},{"t_s":38,"rss_mib":1158.78515625},{"t_s":39,"rss_mib":1158.78515625},{"t_s":39,"rss_mib":1158.79296875},{"t_s":39,"rss_mib":1158.79296875},{"t_s":40,"rss_mib":1158.80078125},{"t_s":40,"rss_mib":1158.80859375},{"t_s":40,"rss_mib":1158.80859375},{"t_s":40,"rss_mib":1158.81640625},{"t_s":41,"rss_mib":1158.81640625},{"t_s":41,"rss_mib":1158.82421875},{"t_s":42,"rss_mib":1158.82421875},{"t_s":42,"rss_mib":1158.82421875},{"t_s":42,"rss_mib":1158.83203125},{"t_s":43,"rss_mib":1158.83203125},{"t_s":43,"rss_mib":1158.83203125},{"t_s":43,"rss_mib":1159.03515625},{"t_s":43,"rss_mib":1159.03515625},{"t_s":44,"rss_mib":1159.05078125},{"t_s":44,"rss_mib":1159.05078125},{"t_s":45,"rss_mib":1159.05078125},{"t_s":45,"rss_mib":1159.05078125},{"t_s":45,"rss_mib":1159.06640625},{"t_s":46,"rss_mib":1159.06640625},{"t_s":46,"rss_mib":1159.07421875},{"t_s":46,"rss_mib":1159.07421875},{"t_s":46,"rss_mib":1159.07421875},{"t_s":47,"rss_mib":1159.07421875},{"t_s":47,"rss_mib":1159.08203125},{"t_s":48,"rss_mib":1159.0859375},{"t_s":48,"rss_mib":1159.0859375},{"t_s":48,"rss_mib":1159.09375},{"t_s":49,"rss_mib":1159.09375},{"t_s":49,"rss_mib":1159.09375},{"t_s":49,"rss_mib":1159.09375},{"t_s":50,"rss_mib":1159.109375},{"t_s":50,"rss_mib":1159.109375},{"t_s":50,"rss_mib":1159.125},{"t_s":50,"rss_mib":1159.125},{"t_s":51,"rss_mib":1159.125},{"t_s":52,"rss_mib":1159.1328125},{"t_s":52,"rss_mib":1159.140625},{"t_s":52,"rss_mib":1159.140625},{"t_s":52,"rss_mib":1159.14453125},{"t_s":53,"rss_mib":1159.14453125},{"t_s":53,"rss_mib":1159.19140625},{"t_s":53,"rss_mib":1159.19140625},{"t_s":53,"rss_mib":1159.19140625},{"t_s":54,"rss_mib":1159.34765625},{"t_s":54,"rss_mib":1159.34765625},{"t_s":55,"rss_mib":1159.34765625},{"t_s":55,"rss_mib":1159.359375},{"t_s":55,"rss_mib":1159.359375},{"t_s":55,"rss_mib":1159.3671875},{"t_s":56,"rss_mib":1159.3671875},{"t_s":56,"rss_mib":1159.3671875},{"t_s":56,"rss_mib":1159.375},{"t_s":56,"rss_mib":1159.375},{"t_s":57,"rss_mib":1159.3828125},{"t_s":57,"rss_mib":1159.3828125},{"t_s":58,"rss_mib":1159.3828125},{"t_s":58,"rss_mib":1159.390625},{"t_s":58,"rss_mib":1159.390625},{"t_s":59,"rss_mib":1159.3984375},{"t_s":59,"rss_mib":1159.40625},{"t_s":59,"rss_mib":1159.40625},{"t_s":59,"rss_mib":1159.4140625},{"t_s":59,"rss_mib":1159.4140625}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.010233258,"cost":6.03410124,"memory":433.987096623,"streaming":0.000002284,"streams_sustained":0.000001799,"throughput":333.769957533},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"perf.frontier.1ms.concurrency":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.p99_us":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.rps":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.samples":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"stream.added_gap_p50_us":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_gap_p99_us":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p50_us":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.added_ttft_p99_us":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_direct_samples":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.gap_gw_samples":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_fps":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_headroom":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.streams_sustained_mock_ceiling":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_direct_samples":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"},"stream.ttft_gw_samples":{"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here"}}},"anthropic":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1/messages","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}}}},"openai":{"configurable":true,"served":true,"egress_config":null,"serve_error":"","cells":{"openai":{"served":true,"reason":null,"status":"","path":"/v1/chat/completions","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":718,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":943,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":971,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":709.3196158019531,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1409.801699716714,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.7330563381485045,"certified":true,"suppressed":false},"cost_window_ok":{"value":24883,"certified":true,"suppressed":false},"cost_window_rps":{"value":4146,"certified":true,"suppressed":false},"cost_threads":{"value":142,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.008399308765020295,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":2688,"certified":true,"suppressed":false},"concurrency":2,"p99_us":944,"first_disqualified_conc":4,"lower_bound":false,"samples":16131},{"bound_ms":5,"rps":{"value":5685,"certified":true,"suppressed":false},"concurrency":8,"p99_us":2498,"first_disqualified_conc":16,"lower_bound":false,"samples":34120},{"bound_ms":10,"rps":{"value":6095,"certified":true,"suppressed":false},"concurrency":32,"p99_us":8960,"first_disqualified_conc":64,"lower_bound":false,"samples":36576},{"bound_ms":50,"rps":{"value":6138,"certified":true,"suppressed":false},"concurrency":64,"p99_us":14447,"first_disqualified_conc":256,"lower_bound":false,"samples":36832},{"bound_ms":100,"rps":{"value":6138,"certified":true,"suppressed":false},"concurrency":64,"p99_us":14447,"first_disqualified_conc":512,"lower_bound":false,"samples":36832},{"bound_ms":null,"rps":{"value":6138,"certified":true,"suppressed":false},"concurrency":64,"p99_us":14447,"first_disqualified_conc":null,"lower_bound":false,"samples":36832}],"sweep":[{"conc":1,"ok":8277,"rps":1379,"p99_us":936,"fail":0},{"conc":1,"ok":8251,"rps":1374,"p99_us":942,"fail":0},{"conc":1,"ok":8197,"rps":1365,"p99_us":953,"fail":0},{"conc":2,"ok":15569,"rps":2594,"p99_us":966,"fail":0},{"conc":2,"ok":15751,"rps":2624,"p99_us":957,"fail":0},{"conc":2,"ok":16131,"rps":2688,"p99_us":944,"fail":0},{"conc":4,"ok":25079,"rps":4179,"p99_us":1611,"fail":0},{"conc":4,"ok":24950,"rps":4157,"p99_us":1591,"fail":0},{"conc":4,"ok":31773,"rps":5294,"p99_us":1068,"fail":0},{"conc":8,"ok":26516,"rps":4418,"p99_us":2587,"fail":0},{"conc":8,"ok":26041,"rps":4339,"p99_us":3958,"fail":0},{"conc":8,"ok":34120,"rps":5685,"p99_us":2498,"fail":0},{"conc":16,"ok":35427,"rps":5903,"p99_us":6656,"fail":0},{"conc":16,"ok":34776,"rps":5795,"p99_us":6635,"fail":0},{"conc":16,"ok":34142,"rps":5689,"p99_us":7386,"fail":0},{"conc":32,"ok":36220,"rps":6036,"p99_us":10493,"fail":0},{"conc":32,"ok":36576,"rps":6095,"p99_us":8960,"fail":0},{"conc":32,"ok":36312,"rps":6051,"p99_us":10425,"fail":0},{"conc":64,"ok":36832,"rps":6138,"p99_us":14447,"fail":0},{"conc":64,"ok":36784,"rps":6130,"p99_us":14497,"fail":0},{"conc":64,"ok":36765,"rps":6126,"p99_us":14741,"fail":0},{"conc":128,"ok":36609,"rps":6100,"p99_us":33886,"fail":0},{"conc":128,"ok":36625,"rps":6102,"p99_us":33878,"fail":0},{"conc":128,"ok":36583,"rps":6095,"p99_us":28695,"fail":0},{"conc":256,"ok":34863,"rps":5809,"p99_us":60903,"fail":0},{"conc":256,"ok":35402,"rps":5899,"p99_us":65344,"fail":0},{"conc":256,"ok":35191,"rps":5864,"p99_us":57589,"fail":0},{"conc":512,"ok":33836,"rps":5638,"p99_us":118651,"fail":0},{"conc":512,"ok":33751,"rps":5624,"p99_us":110609,"fail":0},{"conc":512,"ok":33893,"rps":5648,"p99_us":118801,"fail":0},{"conc":1024,"ok":33606,"rps":5599,"p99_us":1034532,"fail":0},{"conc":1024,"ok":33944,"rps":5656,"p99_us":227956,"fail":0},{"conc":1024,"ok":33912,"rps":5650,"p99_us":249044,"fail":0},{"conc":2048,"ok":33804,"rps":5632,"p99_us":810707,"fail":0},{"conc":2048,"ok":33833,"rps":5637,"p99_us":749523,"fail":0},{"conc":2048,"ok":33997,"rps":5664,"p99_us":855634,"fail":0},{"conc":4096,"ok":33921,"rps":5651,"p99_us":772468,"fail":0},{"conc":4096,"ok":33888,"rps":5647,"p99_us":406404,"fail":0},{"conc":4096,"ok":33990,"rps":5663,"p99_us":708027,"fail":0},{"conc":8192,"ok":34130,"rps":5688,"p99_us":653786,"fail":0},{"conc":8192,"ok":34049,"rps":5673,"p99_us":597661,"fail":0},{"conc":8192,"ok":34215,"rps":5701,"p99_us":809671,"fail":0},{"conc":16384,"ok":34112,"rps":5684,"p99_us":677993,"fail":0},{"conc":16384,"ok":34143,"rps":5690,"p99_us":664790,"fail":0},{"conc":16384,"ok":34041,"rps":5671,"p99_us":761220,"fail":0},{"conc":32768,"ok":26116,"rps":4352,"p99_us":3063275,"fail":0},{"conc":32768,"ok":30806,"rps":5133,"p99_us":2651630,"fail":0},{"conc":32768,"ok":29598,"rps":4932,"p99_us":2602395,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1374,"certified":true,"suppressed":false},"p99_us":942,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":2624,"certified":true,"suppressed":false},"p99_us":957,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":4179,"certified":true,"suppressed":false},"p99_us":1591,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":4418,"certified":true,"suppressed":false},"p99_us":2587,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":5795,"certified":true,"suppressed":false},"p99_us":6656,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":6051,"certified":true,"suppressed":false},"p99_us":10425,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":6130,"certified":true,"suppressed":false},"p99_us":14497,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":6100,"certified":true,"suppressed":false},"p99_us":33878,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":5864,"certified":true,"suppressed":false},"p99_us":60903,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":5638,"certified":true,"suppressed":false},"p99_us":118651,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":5650,"certified":true,"suppressed":false},"p99_us":249044,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":5637,"certified":true,"suppressed":false},"p99_us":810707,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":5651,"certified":true,"suppressed":false},"p99_us":708027,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":5688,"certified":true,"suppressed":false},"p99_us":653786,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":5684,"certified":true,"suppressed":false},"p99_us":677993,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":4932,"certified":true,"suppressed":false},"p99_us":2651630,"windows":3,"clean_windows":3}],"reverify_note":"openai>openai is a same-dialect cell: there is no translation to prove, so a request-shape check upstream cannot distinguish a translating gateway from a forwarding one","c1_note":"the c=1 percentiles are taken over 7996 successful gateway round trip(s) and 262593 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":7996,"direct_c1_samples":262593},"stream":{"stream_served":true,"stream_error":null,"reason":null,"stream_c1_note":"the c=1 streaming legs read 64 frame(s) through the gateway and 64 direct to the mock, out of a 64 frame budget. The added-GAP figures are percentiles over the intervals inside those two streams. The added-TTFT figures are percentiles over separate probes - 100 through the gateway and 100 direct, out of 100 attempted per leg - so they carry that weight and no more","added_ttft_p50_us":{"value":881,"certified":true,"suppressed":false},"added_ttft_p99_us":{"value":1103,"certified":true,"suppressed":false},"added_gap_p50_us":{"value":2,"certified":true,"suppressed":false},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"below_resolution","detail":"the gateway's own inter-frame gap at this percentile (21141us) came in under the mock's (21153us), which a proxy cannot really do - the added gap here is below what this rig can resolve against the mock's 20ms pacing"},"streams_sustained_fps":{"value":52911.767627714544,"certified":true,"suppressed":false,"headroom":0.617121104958904,"rig_ceiling":85739.68253968254},"streams_sustained":{"value":1688,"certified":true,"suppressed":false,"headroom":0.617121104958904,"rig_ceiling":85739.68253968254},"ttft_gw_samples":100,"ttft_direct_samples":100,"gap_gw_samples":63,"gap_direct_samples":63},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1163.541015625,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1170.94140625,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1171.9921875,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1181.8125,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1287.14453125,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":66,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.8963186649029332,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":0.6061112080903336,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1164.1640625},{"t_s":2,"rss_mib":1180.8671875},{"t_s":3,"rss_mib":1180.875},{"t_s":5,"rss_mib":1180.94140625},{"t_s":6,"rss_mib":1180.94140625},{"t_s":7,"rss_mib":1181.12109375},{"t_s":9,"rss_mib":1181.12109375},{"t_s":11,"rss_mib":1181.1328125},{"t_s":12,"rss_mib":1181.1328125},{"t_s":14,"rss_mib":1181.23828125},{"t_s":15,"rss_mib":1181.23828125},{"t_s":16,"rss_mib":1181.2734375},{"t_s":18,"rss_mib":1181.2734375},{"t_s":20,"rss_mib":1181.3359375},{"t_s":21,"rss_mib":1181.3359375},{"t_s":22,"rss_mib":1181.375},{"t_s":24,"rss_mib":1181.375},{"t_s":26,"rss_mib":1181.41796875},{"t_s":26,"rss_mib":1181.41796875},{"t_s":28,"rss_mib":1181.43359375},{"t_s":30,"rss_mib":1181.43359375},{"t_s":32,"rss_mib":1181.453125},{"t_s":33,"rss_mib":1181.46484375},{"t_s":34,"rss_mib":1181.48046875},{"t_s":36,"rss_mib":1181.48046875},{"t_s":37,"rss_mib":1181.50390625},{"t_s":39,"rss_mib":1181.50390625},{"t_s":41,"rss_mib":1181.51171875},{"t_s":42,"rss_mib":1181.51171875},{"t_s":43,"rss_mib":1181.5390625},{"t_s":45,"rss_mib":1181.5390625},{"t_s":47,"rss_mib":1181.64453125},{"t_s":48,"rss_mib":1181.64453125},{"t_s":50,"rss_mib":1181.73046875},{"t_s":51,"rss_mib":1181.73046875},{"t_s":52,"rss_mib":1181.7890625},{"t_s":54,"rss_mib":1181.8125},{"t_s":54,"rss_mib":1178.03125},{"t_s":57,"rss_mib":1178.0390625},{"t_s":59,"rss_mib":1178.05859375},{"t_s":60,"rss_mib":1178.05859375},{"t_s":60,"rss_mib":1178.1796875},{"t_s":63,"rss_mib":1178.1796875},{"t_s":65,"rss_mib":1178.1875},{"t_s":66,"rss_mib":1178.1953125},{"t_s":69,"rss_mib":1178.1953125},{"t_s":69,"rss_mib":1177.453125},{"t_s":72,"rss_mib":1177.453125},{"t_s":75,"rss_mib":1177.45703125},{"t_s":78,"rss_mib":1177.45703125},{"t_s":79,"rss_mib":1177.5},{"t_s":81,"rss_mib":1177.5},{"t_s":83,"rss_mib":1177.51171875},{"t_s":84,"rss_mib":1177.51171875},{"t_s":85,"rss_mib":1177.5546875},{"t_s":87,"rss_mib":1177.5546875},{"t_s":89,"rss_mib":1177.5703125},{"t_s":90,"rss_mib":1177.5703125},{"t_s":91,"rss_mib":1177.578125},{"t_s":93,"rss_mib":1177.578125},{"t_s":94,"rss_mib":1177.58984375},{"t_s":96,"rss_mib":1177.58984375},{"t_s":98,"rss_mib":1177.63671875},{"t_s":99,"rss_mib":1177.63671875},{"t_s":100,"rss_mib":1177.65234375},{"t_s":102,"rss_mib":1177.65234375},{"t_s":102,"rss_mib":1177.66015625},{"t_s":105,"rss_mib":1177.66015625},{"t_s":107,"rss_mib":1177.67578125},{"t_s":108,"rss_mib":1177.67578125},{"t_s":111,"rss_mib":1177.8515625},{"t_s":111,"rss_mib":1177.8515625},{"t_s":113,"rss_mib":1177.94921875},{"t_s":114,"rss_mib":1177.94921875},{"t_s":117,"rss_mib":1177.96484375},{"t_s":117,"rss_mib":1177.96484375},{"t_s":120,"rss_mib":1177.96484375},{"t_s":122,"rss_mib":1177.9765625},{"t_s":123,"rss_mib":1177.9765625},{"t_s":126,"rss_mib":1177.9765625},{"t_s":127,"rss_mib":1178.05078125},{"t_s":129,"rss_mib":1178.05078125},{"t_s":130,"rss_mib":1178.0703125},{"t_s":132,"rss_mib":1178.0703125},{"t_s":134,"rss_mib":1178.12890625},{"t_s":135,"rss_mib":1178.12890625},{"t_s":137,"rss_mib":1178.140625},{"t_s":138,"rss_mib":1178.140625},{"t_s":141,"rss_mib":1178.140625},{"t_s":142,"rss_mib":1178.14453125},{"t_s":144,"rss_mib":1178.14453125},{"t_s":147,"rss_mib":1178.14453125},{"t_s":150,"rss_mib":1178.14453125},{"t_s":152,"rss_mib":1178.15234375},{"t_s":153,"rss_mib":1178.15234375},{"t_s":156,"rss_mib":1178.15234375},{"t_s":159,"rss_mib":1178.15234375},{"t_s":159,"rss_mib":1178.62890625},{"t_s":162,"rss_mib":1178.62890625},{"t_s":163,"rss_mib":1178.7734375},{"t_s":165,"rss_mib":1178.7734375},{"t_s":166,"rss_mib":1178.78515625},{"t_s":168,"rss_mib":1178.78515625},{"t_s":170,"rss_mib":1178.84375},{"t_s":171,"rss_mib":1178.84375},{"t_s":171,"rss_mib":1178.88671875},{"t_s":174,"rss_mib":1178.88671875},{"t_s":174,"rss_mib":1171.26953125},{"t_s":177,"rss_mib":1171.27734375},{"t_s":179,"rss_mib":1170.44921875},{"t_s":180,"rss_mib":1170.44921875},{"t_s":183,"rss_mib":1170.44921875},{"t_s":185,"rss_mib":1170.45703125},{"t_s":186,"rss_mib":1170.45703125},{"t_s":188,"rss_mib":1170.47265625},{"t_s":189,"rss_mib":1170.47265625},{"t_s":191,"rss_mib":1170.49609375},{"t_s":192,"rss_mib":1170.50390625},{"t_s":195,"rss_mib":1170.52734375},{"t_s":195,"rss_mib":1170.52734375},{"t_s":198,"rss_mib":1170.52734375},{"t_s":199,"rss_mib":1170.5390625},{"t_s":201,"rss_mib":1170.5390625},{"t_s":201,"rss_mib":1170.546875},{"t_s":204,"rss_mib":1170.546875},{"t_s":205,"rss_mib":1170.5546875},{"t_s":207,"rss_mib":1170.5546875},{"t_s":210,"rss_mib":1170.5546875},{"t_s":211,"rss_mib":1170.5703125},{"t_s":213,"rss_mib":1170.5703125},{"t_s":216,"rss_mib":1170.5703125},{"t_s":218,"rss_mib":1170.57421875},{"t_s":219,"rss_mib":1170.57421875},{"t_s":222,"rss_mib":1170.57421875},{"t_s":224,"rss_mib":1170.58984375},{"t_s":225,"rss_mib":1170.58984375},{"t_s":227,"rss_mib":1170.60546875},{"t_s":228,"rss_mib":1170.60546875},{"t_s":230,"rss_mib":1170.66015625},{"t_s":231,"rss_mib":1170.73046875},{"t_s":234,"rss_mib":1170.7734375},{"t_s":234,"rss_mib":1170.7734375},{"t_s":236,"rss_mib":1170.7890625},{"t_s":237,"rss_mib":1170.7890625},{"t_s":239,"rss_mib":1170.80078125},{"t_s":240,"rss_mib":1170.80078125},{"t_s":241,"rss_mib":1170.80859375},{"t_s":243,"rss_mib":1170.80859375},{"t_s":245,"rss_mib":1170.88671875},{"t_s":246,"rss_mib":1170.88671875},{"t_s":249,"rss_mib":1170.9140625},{"t_s":249,"rss_mib":1170.9140625},{"t_s":252,"rss_mib":1170.93359375},{"t_s":252,"rss_mib":1170.93359375},{"t_s":254,"rss_mib":1170.94140625},{"t_s":255,"rss_mib":1170.94140625},{"t_s":258,"rss_mib":1170.94140625},{"t_s":258,"rss_mib":1170.94921875},{"t_s":261,"rss_mib":1170.94921875},{"t_s":264,"rss_mib":1170.94921875},{"t_s":264,"rss_mib":1170.95703125},{"t_s":267,"rss_mib":1170.95703125},{"t_s":269,"rss_mib":1170.96875},{"t_s":270,"rss_mib":1170.96875},{"t_s":272,"rss_mib":1171.10546875},{"t_s":273,"rss_mib":1171.10546875},{"t_s":273,"rss_mib":1171.11328125},{"t_s":276,"rss_mib":1171.11328125},{"t_s":277,"rss_mib":1171.1171875},{"t_s":279,"rss_mib":1171.1171875},{"t_s":279,"rss_mib":1171.125},{"t_s":282,"rss_mib":1171.125},{"t_s":284,"rss_mib":1171.20703125},{"t_s":285,"rss_mib":1171.20703125},{"t_s":287,"rss_mib":1171.39453125},{"t_s":288,"rss_mib":1171.55859375},{"t_s":291,"rss_mib":1171.69140625},{"t_s":291,"rss_mib":1171.69140625},{"t_s":294,"rss_mib":1171.70703125},{"t_s":294,"rss_mib":1171.71875},{"t_s":294,"rss_mib":1171.66015625},{"t_s":297,"rss_mib":1171.66015625},{"t_s":299,"rss_mib":1171.66796875},{"t_s":300,"rss_mib":1171.66796875},{"t_s":303,"rss_mib":1171.66796875},{"t_s":306,"rss_mib":1171.66796875},{"t_s":309,"rss_mib":1171.66796875},{"t_s":309,"rss_mib":1171.671875},{"t_s":312,"rss_mib":1171.671875},{"t_s":313,"rss_mib":1171.6796875},{"t_s":315,"rss_mib":1171.7109375},{"t_s":316,"rss_mib":1171.7421875},{"t_s":318,"rss_mib":1171.7421875},{"t_s":321,"rss_mib":1171.74609375},{"t_s":321,"rss_mib":1171.79296875},{"t_s":324,"rss_mib":1171.79296875},{"t_s":325,"rss_mib":1171.80078125},{"t_s":327,"rss_mib":1171.80078125},{"t_s":329,"rss_mib":1171.95703125},{"t_s":330,"rss_mib":1171.95703125},{"t_s":333,"rss_mib":1171.95703125},{"t_s":335,"rss_mib":1171.96875},{"t_s":336,"rss_mib":1171.96875},{"t_s":337,"rss_mib":1171.98046875},{"t_s":339,"rss_mib":1171.98046875},{"t_s":339,"rss_mib":1171.9921875},{"t_s":342,"rss_mib":1171.9921875},{"t_s":345,"rss_mib":1171.9921875},{"t_s":348,"rss_mib":1176.25},{"t_s":348,"rss_mib":1172.86328125},{"t_s":351,"rss_mib":1172.94140625},{"t_s":354,"rss_mib":1172.94140625},{"t_s":357,"rss_mib":1172.94140625},{"t_s":359,"rss_mib":1172.94921875},{"t_s":360,"rss_mib":1172.94921875}],"idle_rss_series":[{"t_s":0,"rss_mib":1163.33203125},{"t_s":0,"rss_mib":1163.41015625},{"t_s":0,"rss_mib":1163.41015625},{"t_s":0,"rss_mib":1163.41015625},{"t_s":1,"rss_mib":1163.41015625},{"t_s":1,"rss_mib":1163.41015625},{"t_s":2,"rss_mib":1163.41015625},{"t_s":3,"rss_mib":1163.41015625},{"t_s":3,"rss_mib":1163.41015625},{"t_s":3,"rss_mib":1163.41015625},{"t_s":4,"rss_mib":1163.4140625},{"t_s":4,"rss_mib":1163.41796875},{"t_s":4,"rss_mib":1163.421875},{"t_s":4,"rss_mib":1163.421875},{"t_s":5,"rss_mib":1163.421875},{"t_s":5,"rss_mib":1163.421875},{"t_s":6,"rss_mib":1163.421875},{"t_s":6,"rss_mib":1163.421875},{"t_s":7,"rss_mib":1163.421875},{"t_s":7,"rss_mib":1163.421875},{"t_s":8,"rss_mib":1163.421875},{"t_s":8,"rss_mib":1163.421875},{"t_s":9,"rss_mib":1163.421875},{"t_s":9,"rss_mib":1163.4453125},{"t_s":10,"rss_mib":1163.4453125},{"t_s":10,"rss_mib":1163.4453125},{"t_s":10,"rss_mib":1163.4453125},{"t_s":11,"rss_mib":1163.4453125},{"t_s":11,"rss_mib":1163.44921875},{"t_s":12,"rss_mib":1163.44921875},{"t_s":13,"rss_mib":1163.44921875},{"t_s":13,"rss_mib":1163.44921875},{"t_s":13,"rss_mib":1163.44921875},{"t_s":14,"rss_mib":1163.44921875},{"t_s":14,"rss_mib":1163.44921875},{"t_s":15,"rss_mib":1163.44921875},{"t_s":16,"rss_mib":1163.44921875},{"t_s":16,"rss_mib":1163.44921875},{"t_s":16,"rss_mib":1163.44921875},{"t_s":17,"rss_mib":1163.44921875},{"t_s":17,"rss_mib":1163.44921875},{"t_s":18,"rss_mib":1163.44921875},{"t_s":18,"rss_mib":1163.44921875},{"t_s":19,"rss_mib":1163.44921875},{"t_s":20,"rss_mib":1163.45703125},{"t_s":20,"rss_mib":1163.45703125},{"t_s":20,"rss_mib":1163.45703125},{"t_s":21,"rss_mib":1163.46484375},{"t_s":21,"rss_mib":1163.47265625},{"t_s":21,"rss_mib":1163.4765625},{"t_s":22,"rss_mib":1163.4765625},{"t_s":23,"rss_mib":1163.4765625},{"t_s":23,"rss_mib":1163.4765625},{"t_s":23,"rss_mib":1163.484375},{"t_s":23,"rss_mib":1163.484375},{"t_s":24,"rss_mib":1163.51171875},{"t_s":24,"rss_mib":1163.51171875},{"t_s":24,"rss_mib":1163.51171875},{"t_s":25,"rss_mib":1163.515625},{"t_s":25,"rss_mib":1163.515625},{"t_s":25,"rss_mib":1163.5234375},{"t_s":26,"rss_mib":1163.5234375},{"t_s":26,"rss_mib":1163.5234375},{"t_s":26,"rss_mib":1163.5234375},{"t_s":27,"rss_mib":1163.5234375},{"t_s":27,"rss_mib":1163.5234375},{"t_s":28,"rss_mib":1163.5234375},{"t_s":28,"rss_mib":1163.53125},{"t_s":29,"rss_mib":1163.53515625},{"t_s":29,"rss_mib":1163.53515625},{"t_s":29,"rss_mib":1163.5390625},{"t_s":30,"rss_mib":1163.54296875},{"t_s":30,"rss_mib":1163.54296875},{"t_s":30,"rss_mib":1163.54296875},{"t_s":31,"rss_mib":1163.55078125},{"t_s":31,"rss_mib":1163.55859375},{"t_s":32,"rss_mib":1163.55859375},{"t_s":32,"rss_mib":1163.55859375},{"t_s":33,"rss_mib":1163.55859375},{"t_s":33,"rss_mib":1163.55859375},{"t_s":33,"rss_mib":1163.56640625},{"t_s":33,"rss_mib":1163.56640625},{"t_s":34,"rss_mib":1163.56640625},{"t_s":34,"rss_mib":1163.56640625},{"t_s":35,"rss_mib":1163.56640625},{"t_s":36,"rss_mib":1163.57421875},{"t_s":36,"rss_mib":1163.57421875},{"t_s":36,"rss_mib":1163.57421875},{"t_s":37,"rss_mib":1163.57421875},{"t_s":37,"rss_mib":1163.57421875},{"t_s":38,"rss_mib":1163.57421875},{"t_s":39,"rss_mib":1163.578125},{"t_s":39,"rss_mib":1163.58203125},{"t_s":39,"rss_mib":1163.58203125},{"t_s":39,"rss_mib":1163.58984375},{"t_s":40,"rss_mib":1163.58984375},{"t_s":40,"rss_mib":1163.58984375},{"t_s":40,"rss_mib":1163.58984375},{"t_s":41,"rss_mib":1163.58984375},{"t_s":41,"rss_mib":1163.59765625},{"t_s":42,"rss_mib":1163.60546875},{"t_s":42,"rss_mib":1163.60546875},{"t_s":43,"rss_mib":1163.60546875},{"t_s":43,"rss_mib":1163.60546875},{"t_s":43,"rss_mib":1163.60546875},{"t_s":44,"rss_mib":1163.61328125},{"t_s":44,"rss_mib":1163.61328125},{"t_s":44,"rss_mib":1163.6171875},{"t_s":45,"rss_mib":1163.6171875},{"t_s":46,"rss_mib":1163.6171875},{"t_s":46,"rss_mib":1163.62890625},{"t_s":46,"rss_mib":1163.63671875},{"t_s":46,"rss_mib":1163.63671875},{"t_s":47,"rss_mib":1163.63671875},{"t_s":47,"rss_mib":1163.63671875},{"t_s":48,"rss_mib":1163.78125},{"t_s":48,"rss_mib":1163.78125},{"t_s":48,"rss_mib":1163.7890625},{"t_s":49,"rss_mib":1163.7890625},{"t_s":49,"rss_mib":1163.796875},{"t_s":49,"rss_mib":1163.796875},{"t_s":49,"rss_mib":1163.80859375},{"t_s":49,"rss_mib":1163.80859375},{"t_s":50,"rss_mib":1163.84765625},{"t_s":50,"rss_mib":1163.84765625},{"t_s":50,"rss_mib":1163.85546875},{"t_s":50,"rss_mib":1163.85546875},{"t_s":51,"rss_mib":1163.94921875},{"t_s":51,"rss_mib":1163.94921875},{"t_s":51,"rss_mib":1163.95703125},{"t_s":52,"rss_mib":1163.96484375},{"t_s":52,"rss_mib":1163.96484375},{"t_s":53,"rss_mib":1163.96484375},{"t_s":53,"rss_mib":1163.97265625},{"t_s":53,"rss_mib":1163.98046875},{"t_s":53,"rss_mib":1163.984375},{"t_s":54,"rss_mib":1164.015625},{"t_s":54,"rss_mib":1164.015625},{"t_s":54,"rss_mib":1164.01953125},{"t_s":55,"rss_mib":1164.01953125},{"t_s":55,"rss_mib":1164.02734375},{"t_s":55,"rss_mib":1164.02734375},{"t_s":55,"rss_mib":1164.03515625},{"t_s":56,"rss_mib":1164.0390625},{"t_s":56,"rss_mib":1164.0390625},{"t_s":56,"rss_mib":1164.0390625},{"t_s":57,"rss_mib":1164.0390625},{"t_s":57,"rss_mib":1164.0546875},{"t_s":57,"rss_mib":1164.05859375},{"t_s":58,"rss_mib":1164.05859375},{"t_s":59,"rss_mib":1164.05859375},{"t_s":59,"rss_mib":1164.05859375},{"t_s":59,"rss_mib":1164.06640625},{"t_s":59,"rss_mib":1164.06640625}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.009584148,"cost":6.034291549,"memory":436.869105164,"streaming":2.727068519,"streams_sustained":163.276888075,"throughput":323.054589636},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"stream.added_gap_p99_us":{"reason":"below_resolution","detail":"the gateway's own inter-frame gap at this percentile (21141us) came in under the mock's (21153us), which a proxy cannot really do - the added gap here is below what this rig can resolve against the mock's 20ms pacing"}}},"gemini":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1beta/models/gpt-4o-mini:generateContent","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"bedrock":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/model/gpt-4o-mini/converse","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"cohere":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v2/chat","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}},"anthropic":{"served":true,"reason":null,"status":"","path":"/v1/messages","verdict_note":"","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":{"added_latency_p50_us":{"value":954,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":1252,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":1280,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":968.311271560369,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1032.7257663628832,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.5011834686494262,"certified":true,"suppressed":false},"cost_window_ok":{"value":12465,"certified":true,"suppressed":false},"cost_window_rps":{"value":2076,"certified":true,"suppressed":false},"cost_threads":{"value":142,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.00008022462896109105,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":3867,"certified":true,"suppressed":false},"concurrency":8,"p99_us":4852,"first_disqualified_conc":16,"lower_bound":false,"samples":23211},{"bound_ms":10,"rps":{"value":3999,"certified":true,"suppressed":false},"concurrency":16,"p99_us":5892,"first_disqualified_conc":32,"lower_bound":false,"samples":23998},{"bound_ms":50,"rps":{"value":4118,"certified":true,"suppressed":false},"concurrency":64,"p99_us":23767,"first_disqualified_conc":256,"lower_bound":false,"samples":24713},{"bound_ms":100,"rps":{"value":4118,"certified":true,"suppressed":false},"concurrency":64,"p99_us":23767,"first_disqualified_conc":512,"lower_bound":false,"samples":24713},{"bound_ms":null,"rps":{"value":4118,"certified":true,"suppressed":false},"concurrency":64,"p99_us":23767,"first_disqualified_conc":null,"lower_bound":false,"samples":24713}],"sweep":[{"conc":1,"ok":5717,"rps":951,"p99_us":1346,"fail":0},{"conc":1,"ok":5738,"rps":956,"p99_us":1332,"fail":0},{"conc":1,"ok":5753,"rps":957,"p99_us":1322,"fail":0},{"conc":2,"ok":6020,"rps":1003,"p99_us":2461,"fail":0},{"conc":2,"ok":11272,"rps":1878,"p99_us":1363,"fail":0},{"conc":2,"ok":6141,"rps":1023,"p99_us":2375,"fail":0},{"conc":4,"ok":17326,"rps":2887,"p99_us":2475,"fail":0},{"conc":4,"ok":17300,"rps":2882,"p99_us":2443,"fail":0},{"conc":4,"ok":17357,"rps":2892,"p99_us":2465,"fail":0},{"conc":8,"ok":18113,"rps":3018,"p99_us":4859,"fail":0},{"conc":8,"ok":23161,"rps":3859,"p99_us":3875,"fail":0},{"conc":8,"ok":23211,"rps":3867,"p99_us":4852,"fail":0},{"conc":16,"ok":23787,"rps":3964,"p99_us":8534,"fail":0},{"conc":16,"ok":23998,"rps":3999,"p99_us":5892,"fail":0},{"conc":16,"ok":23780,"rps":3962,"p99_us":8573,"fail":0},{"conc":32,"ok":24506,"rps":4083,"p99_us":16231,"fail":0},{"conc":32,"ok":24466,"rps":4077,"p99_us":15467,"fail":0},{"conc":32,"ok":24518,"rps":4085,"p99_us":12056,"fail":0},{"conc":64,"ok":24706,"rps":4117,"p99_us":21224,"fail":0},{"conc":64,"ok":24669,"rps":4110,"p99_us":24591,"fail":0},{"conc":64,"ok":24713,"rps":4118,"p99_us":23767,"fail":0},{"conc":128,"ok":24526,"rps":4087,"p99_us":44306,"fail":0},{"conc":128,"ok":24439,"rps":4073,"p99_us":41656,"fail":0},{"conc":128,"ok":24493,"rps":4081,"p99_us":41564,"fail":0},{"conc":256,"ok":23540,"rps":3921,"p99_us":85227,"fail":0},{"conc":256,"ok":23624,"rps":3936,"p99_us":87778,"fail":0},{"conc":256,"ok":23574,"rps":3928,"p99_us":97984,"fail":0},{"conc":512,"ok":23210,"rps":3867,"p99_us":168272,"fail":0},{"conc":512,"ok":23285,"rps":3879,"p99_us":173506,"fail":0},{"conc":512,"ok":23258,"rps":3874,"p99_us":171326,"fail":0},{"conc":1024,"ok":23373,"rps":3894,"p99_us":1268520,"fail":0},{"conc":1024,"ok":23422,"rps":3902,"p99_us":785073,"fail":0},{"conc":1024,"ok":23347,"rps":3890,"p99_us":1186898,"fail":0},{"conc":2048,"ok":23011,"rps":3834,"p99_us":1111092,"fail":0},{"conc":2048,"ok":23050,"rps":3841,"p99_us":1304169,"fail":0},{"conc":2048,"ok":23040,"rps":3839,"p99_us":1136848,"fail":0},{"conc":4096,"ok":22936,"rps":3821,"p99_us":1126822,"fail":0},{"conc":4096,"ok":23069,"rps":3844,"p99_us":1165763,"fail":0},{"conc":4096,"ok":22895,"rps":3814,"p99_us":1118164,"fail":0},{"conc":8192,"ok":23001,"rps":3832,"p99_us":915736,"fail":0},{"conc":8192,"ok":23095,"rps":3848,"p99_us":1201301,"fail":0},{"conc":8192,"ok":23091,"rps":3848,"p99_us":1206709,"fail":0},{"conc":16384,"ok":23064,"rps":3843,"p99_us":1055941,"fail":0},{"conc":16384,"ok":23064,"rps":3843,"p99_us":919993,"fail":0},{"conc":16384,"ok":22917,"rps":3819,"p99_us":1028231,"fail":0},{"conc":32768,"ok":20349,"rps":3390,"p99_us":2908493,"fail":0},{"conc":32768,"ok":22058,"rps":3675,"p99_us":2580108,"fail":0},{"conc":32768,"ok":19591,"rps":3264,"p99_us":2594305,"fail":0}],"rungs":[{"conc":1,"rps":{"value":956,"certified":true,"suppressed":false},"p99_us":1332,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":1023,"certified":true,"suppressed":false},"p99_us":2375,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":2887,"certified":true,"suppressed":false},"p99_us":2465,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":3859,"certified":true,"suppressed":false},"p99_us":4852,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":3964,"certified":true,"suppressed":false},"p99_us":8534,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":4083,"certified":true,"suppressed":false},"p99_us":15467,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":4117,"certified":true,"suppressed":false},"p99_us":23767,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":4081,"certified":true,"suppressed":false},"p99_us":41656,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":3928,"certified":true,"suppressed":false},"p99_us":87778,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":3874,"certified":true,"suppressed":false},"p99_us":171326,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":3894,"certified":true,"suppressed":false},"p99_us":1186898,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":3839,"certified":true,"suppressed":false},"p99_us":1136848,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":3821,"certified":true,"suppressed":false},"p99_us":1126822,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":3848,"certified":true,"suppressed":false},"p99_us":1201301,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":3843,"certified":true,"suppressed":false},"p99_us":1028231,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":3390,"certified":true,"suppressed":false},"p99_us":2594305,"windows":3,"clean_windows":3}],"egress_reverified":true,"c1_note":"the c=1 percentiles are taken over 6058 successful gateway round trip(s) and 259722 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":6058,"direct_c1_samples":259722},"stream":{"stream_served":true,"stream_error":null,"reason":null,"stream_c1_note":"the c=1 streaming legs read 64 frame(s) through the gateway and 64 direct to the mock, out of a 64 frame budget. The added-GAP figures are percentiles over the intervals inside those two streams. The added-TTFT figures are percentiles over separate probes - 100 through the gateway and 100 direct, out of 100 attempted per leg - so they carry that weight and no more","added_ttft_p50_us":{"value":1262,"certified":true,"suppressed":false},"added_ttft_p99_us":{"value":1705,"certified":true,"suppressed":false},"added_gap_p50_us":{"value":8,"certified":true,"suppressed":false},"added_gap_p99_us":{"value":163,"certified":true,"suppressed":false},"streams_sustained_fps":{"value":19376.729819513093,"certified":true,"suppressed":false,"headroom":0.07497629094372328,"rig_ceiling":258438.09523809524},"streams_sustained":{"value":5088,"certified":true,"suppressed":false,"headroom":0.07497629094372328,"rig_ceiling":258438.09523809524},"ttft_gw_samples":100,"ttft_direct_samples":100,"gap_gw_samples":63,"gap_direct_samples":63},"memory":{"protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","served":true,"serve_error":"","load_recipe":null,"idle_rss_mib":{"value":1151.65234375,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":1160.24609375,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":1161.1953125,"certified":true,"suppressed":false},"peak_rss_mib":{"value":1177.22265625,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":1280.44921875,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.5290055438736946,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"idle_static":{"value":1,"certified":true,"suppressed":false},"idle_growth_rate_mib_per_min":{"value":1.003040818815334,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":1152.4453125},{"t_s":2,"rss_mib":1169.20703125},{"t_s":3,"rss_mib":1169.21484375},{"t_s":5,"rss_mib":1169.25390625},{"t_s":6,"rss_mib":1169.27734375},{"t_s":8,"rss_mib":1169.80078125},{"t_s":9,"rss_mib":1169.80859375},{"t_s":11,"rss_mib":1169.83203125},{"t_s":12,"rss_mib":1169.83203125},{"t_s":14,"rss_mib":1169.87109375},{"t_s":15,"rss_mib":1169.87890625},{"t_s":17,"rss_mib":1169.921875},{"t_s":18,"rss_mib":1169.921875},{"t_s":20,"rss_mib":1169.9765625},{"t_s":21,"rss_mib":1169.984375},{"t_s":23,"rss_mib":1170.015625},{"t_s":24,"rss_mib":1170.015625},{"t_s":26,"rss_mib":1170.1171875},{"t_s":27,"rss_mib":1170.1171875},{"t_s":28,"rss_mib":1170.140625},{"t_s":30,"rss_mib":1170.1484375},{"t_s":32,"rss_mib":1170.17578125},{"t_s":33,"rss_mib":1170.18359375},{"t_s":35,"rss_mib":1170.2109375},{"t_s":36,"rss_mib":1170.2109375},{"t_s":38,"rss_mib":1170.28515625},{"t_s":39,"rss_mib":1170.28515625},{"t_s":40,"rss_mib":1170.3046875},{"t_s":42,"rss_mib":1170.3046875},{"t_s":44,"rss_mib":1170.33984375},{"t_s":45,"rss_mib":1170.34765625},{"t_s":47,"rss_mib":1170.46484375},{"t_s":48,"rss_mib":1170.46484375},{"t_s":48,"rss_mib":1170.30859375},{"t_s":51,"rss_mib":1170.46484375},{"t_s":52,"rss_mib":1170.54296875},{"t_s":54,"rss_mib":1170.54296875},{"t_s":56,"rss_mib":1170.71875},{"t_s":57,"rss_mib":1170.71875},{"t_s":59,"rss_mib":1170.81640625},{"t_s":60,"rss_mib":1170.81640625},{"t_s":61,"rss_mib":1170.85546875},{"t_s":63,"rss_mib":1170.85546875},{"t_s":66,"rss_mib":1170.85546875},{"t_s":69,"rss_mib":1170.85546875},{"t_s":69,"rss_mib":1170.6484375},{"t_s":72,"rss_mib":1170.65234375},{"t_s":74,"rss_mib":1170.66015625},{"t_s":75,"rss_mib":1170.66015625},{"t_s":78,"rss_mib":1170.66015625},{"t_s":80,"rss_mib":1170.7421875},{"t_s":81,"rss_mib":1170.7421875},{"t_s":83,"rss_mib":1170.75390625},{"t_s":84,"rss_mib":1170.75390625},{"t_s":86,"rss_mib":1170.80078125},{"t_s":87,"rss_mib":1170.80078125},{"t_s":89,"rss_mib":1170.8203125},{"t_s":90,"rss_mib":1170.8203125},{"t_s":91,"rss_mib":1170.8359375},{"t_s":93,"rss_mib":1170.8359375},{"t_s":94,"rss_mib":1170.8515625},{"t_s":96,"rss_mib":1170.8515625},{"t_s":99,"rss_mib":1170.8515625},{"t_s":100,"rss_mib":1170.86328125},{"t_s":102,"rss_mib":1170.86328125},{"t_s":105,"rss_mib":1170.86328125},{"t_s":107,"rss_mib":1170.87109375},{"t_s":108,"rss_mib":1170.87109375},{"t_s":110,"rss_mib":1171.12109375},{"t_s":111,"rss_mib":1171.12109375},{"t_s":113,"rss_mib":1171.21484375},{"t_s":114,"rss_mib":1171.21484375},{"t_s":114,"rss_mib":1171.26171875},{"t_s":117,"rss_mib":1171.26171875},{"t_s":119,"rss_mib":1171.28125},{"t_s":120,"rss_mib":1171.28125},{"t_s":122,"rss_mib":1171.2890625},{"t_s":123,"rss_mib":1171.2890625},{"t_s":124,"rss_mib":1171.29296875},{"t_s":126,"rss_mib":1171.36328125},{"t_s":128,"rss_mib":1171.37109375},{"t_s":129,"rss_mib":1171.3828125},{"t_s":131,"rss_mib":1171.41015625},{"t_s":132,"rss_mib":1171.4140625},{"t_s":134,"rss_mib":1171.453125},{"t_s":135,"rss_mib":1171.453125},{"t_s":138,"rss_mib":1171.4609375},{"t_s":138,"rss_mib":1171.56640625},{"t_s":141,"rss_mib":1171.56640625},{"t_s":144,"rss_mib":1171.5703125},{"t_s":147,"rss_mib":1171.5703125},{"t_s":150,"rss_mib":1171.5703125},{"t_s":150,"rss_mib":1171.57421875},{"t_s":153,"rss_mib":1171.578125},{"t_s":156,"rss_mib":1171.578125},{"t_s":156,"rss_mib":1171.5859375},{"t_s":159,"rss_mib":1171.5859375},{"t_s":160,"rss_mib":1171.58984375},{"t_s":162,"rss_mib":1171.58984375},{"t_s":165,"rss_mib":1171.58984375},{"t_s":167,"rss_mib":1166.953125},{"t_s":168,"rss_mib":1166.953125},{"t_s":170,"rss_mib":1177.22265625},{"t_s":171,"rss_mib":1167.171875},{"t_s":172,"rss_mib":1167.2421875},{"t_s":174,"rss_mib":1167.2421875},{"t_s":175,"rss_mib":1161.76953125},{"t_s":177,"rss_mib":1161.78125},{"t_s":179,"rss_mib":1159.87109375},{"t_s":180,"rss_mib":1159.87109375},{"t_s":183,"rss_mib":1159.87109375},{"t_s":185,"rss_mib":1159.8828125},{"t_s":186,"rss_mib":1159.8828125},{"t_s":186,"rss_mib":1159.88671875},{"t_s":189,"rss_mib":1159.88671875},{"t_s":190,"rss_mib":1159.91796875},{"t_s":192,"rss_mib":1159.91796875},{"t_s":195,"rss_mib":1159.91796875},{"t_s":198,"rss_mib":1159.91796875},{"t_s":200,"rss_mib":1159.921875},{"t_s":201,"rss_mib":1159.921875},{"t_s":204,"rss_mib":1159.921875},{"t_s":205,"rss_mib":1159.9375},{"t_s":207,"rss_mib":1159.9375},{"t_s":210,"rss_mib":1159.9375},{"t_s":213,"rss_mib":1159.9375},{"t_s":216,"rss_mib":1159.9375},{"t_s":219,"rss_mib":1159.94921875},{"t_s":219,"rss_mib":1159.94921875},{"t_s":220,"rss_mib":1159.953125},{"t_s":222,"rss_mib":1159.953125},{"t_s":225,"rss_mib":1159.953125},{"t_s":228,"rss_mib":1159.95703125},{"t_s":228,"rss_mib":1159.95703125},{"t_s":230,"rss_mib":1160.15625},{"t_s":231,"rss_mib":1160.15625},{"t_s":234,"rss_mib":1160.23046875},{"t_s":234,"rss_mib":1160.23046875},{"t_s":237,"rss_mib":1160.2421875},{"t_s":237,"rss_mib":1160.2421875},{"t_s":240,"rss_mib":1160.2421875},{"t_s":243,"rss_mib":1160.2421875},{"t_s":246,"rss_mib":1160.2421875},{"t_s":249,"rss_mib":1160.2421875},{"t_s":250,"rss_mib":1160.24609375},{"t_s":252,"rss_mib":1160.24609375},{"t_s":255,"rss_mib":1160.24609375},{"t_s":257,"rss_mib":1160.2578125},{"t_s":258,"rss_mib":1160.2578125},{"t_s":260,"rss_mib":1160.26171875},{"t_s":261,"rss_mib":1160.26171875},{"t_s":264,"rss_mib":1160.26171875},{"t_s":264,"rss_mib":1160.328125},{"t_s":267,"rss_mib":1160.328125},{"t_s":270,"rss_mib":1160.328125},{"t_s":270,"rss_mib":1160.33203125},{"t_s":273,"rss_mib":1160.33203125},{"t_s":274,"rss_mib":1160.33984375},{"t_s":276,"rss_mib":1160.33984375},{"t_s":278,"rss_mib":1160.35546875},{"t_s":279,"rss_mib":1160.35546875},{"t_s":280,"rss_mib":1160.359375},{"t_s":282,"rss_mib":1160.359375},{"t_s":285,"rss_mib":1160.359375},{"t_s":287,"rss_mib":1160.421875},{"t_s":288,"rss_mib":1160.421875},{"t_s":290,"rss_mib":1160.5546875},{"t_s":291,"rss_mib":1160.5546875},{"t_s":294,"rss_mib":1160.828125},{"t_s":294,"rss_mib":1160.828125},{"t_s":294,"rss_mib":1160.8359375},{"t_s":297,"rss_mib":1160.8359375},{"t_s":299,"rss_mib":1160.84765625},{"t_s":300,"rss_mib":1160.84765625},{"t_s":301,"rss_mib":1161.0625},{"t_s":303,"rss_mib":1161.0625},{"t_s":304,"rss_mib":1161.0703125},{"t_s":306,"rss_mib":1161.0703125},{"t_s":309,"rss_mib":1161.0703125},{"t_s":310,"rss_mib":1161.08203125},{"t_s":312,"rss_mib":1161.08203125},{"t_s":313,"rss_mib":1161.0859375},{"t_s":315,"rss_mib":1161.0859375},{"t_s":317,"rss_mib":1161.1484375},{"t_s":318,"rss_mib":1161.1484375},{"t_s":321,"rss_mib":1161.1484375},{"t_s":324,"rss_mib":1161.1484375},{"t_s":325,"rss_mib":1161.16796875},{"t_s":327,"rss_mib":1161.16796875},{"t_s":328,"rss_mib":1161.18359375},{"t_s":330,"rss_mib":1161.18359375},{"t_s":333,"rss_mib":1161.18359375},{"t_s":336,"rss_mib":1161.18359375},{"t_s":338,"rss_mib":1161.1875},{"t_s":339,"rss_mib":1161.1875},{"t_s":340,"rss_mib":1161.19140625},{"t_s":342,"rss_mib":1161.19140625},{"t_s":344,"rss_mib":1161.1953125},{"t_s":345,"rss_mib":1161.1953125},{"t_s":348,"rss_mib":1161.1953125},{"t_s":350,"rss_mib":1161.22265625},{"t_s":351,"rss_mib":1161.22265625},{"t_s":352,"rss_mib":1161.2890625},{"t_s":354,"rss_mib":1161.2890625},{"t_s":356,"rss_mib":1161.31640625},{"t_s":357,"rss_mib":1161.31640625},{"t_s":360,"rss_mib":1161.31640625}],"idle_rss_series":[{"t_s":0,"rss_mib":1151.09765625},{"t_s":0,"rss_mib":1151.19921875},{"t_s":0,"rss_mib":1151.19921875},{"t_s":0,"rss_mib":1151.296875},{"t_s":1,"rss_mib":1151.3046875},{"t_s":1,"rss_mib":1151.3046875},{"t_s":2,"rss_mib":1151.3125},{"t_s":2,"rss_mib":1151.3125},{"t_s":2,"rss_mib":1151.32421875},{"t_s":2,"rss_mib":1151.32421875},{"t_s":3,"rss_mib":1151.33203125},{"t_s":3,"rss_mib":1151.33203125},{"t_s":3,"rss_mib":1151.33984375},{"t_s":4,"rss_mib":1151.34765625},{"t_s":4,"rss_mib":1151.34765625},{"t_s":4,"rss_mib":1151.35546875},{"t_s":5,"rss_mib":1151.37109375},{"t_s":5,"rss_mib":1151.37109375},{"t_s":5,"rss_mib":1151.37109375},{"t_s":6,"rss_mib":1151.37890625},{"t_s":6,"rss_mib":1151.38671875},{"t_s":6,"rss_mib":1151.38671875},{"t_s":7,"rss_mib":1151.38671875},{"t_s":7,"rss_mib":1151.38671875},{"t_s":8,"rss_mib":1151.40234375},{"t_s":8,"rss_mib":1151.40234375},{"t_s":8,"rss_mib":1151.41015625},{"t_s":8,"rss_mib":1151.41015625},{"t_s":9,"rss_mib":1151.41015625},{"t_s":10,"rss_mib":1151.41015625},{"t_s":10,"rss_mib":1151.41796875},{"t_s":10,"rss_mib":1151.41796875},{"t_s":10,"rss_mib":1151.41796875},{"t_s":11,"rss_mib":1151.42578125},{"t_s":11,"rss_mib":1151.43359375},{"t_s":11,"rss_mib":1151.43359375},{"t_s":12,"rss_mib":1151.43359375},{"t_s":12,"rss_mib":1151.44140625},{"t_s":13,"rss_mib":1151.44140625},{"t_s":13,"rss_mib":1151.44921875},{"t_s":13,"rss_mib":1151.44921875},{"t_s":13,"rss_mib":1151.44921875},{"t_s":14,"rss_mib":1151.45703125},{"t_s":14,"rss_mib":1151.45703125},{"t_s":14,"rss_mib":1151.45703125},{"t_s":15,"rss_mib":1151.46484375},{"t_s":15,"rss_mib":1151.46484375},{"t_s":15,"rss_mib":1151.47265625},{"t_s":16,"rss_mib":1151.47265625},{"t_s":16,"rss_mib":1151.47265625},{"t_s":16,"rss_mib":1151.48046875},{"t_s":17,"rss_mib":1151.48046875},{"t_s":17,"rss_mib":1151.48828125},{"t_s":17,"rss_mib":1151.48828125},{"t_s":18,"rss_mib":1151.49609375},{"t_s":19,"rss_mib":1151.49609375},{"t_s":19,"rss_mib":1151.5078125},{"t_s":19,"rss_mib":1151.5078125},{"t_s":20,"rss_mib":1151.5078125},{"t_s":20,"rss_mib":1151.515625},{"t_s":20,"rss_mib":1151.515625},{"t_s":20,"rss_mib":1151.515625},{"t_s":21,"rss_mib":1151.5234375},{"t_s":21,"rss_mib":1151.5234375},{"t_s":21,"rss_mib":1151.53125},{"t_s":21,"rss_mib":1151.53125},{"t_s":22,"rss_mib":1151.53125},{"t_s":22,"rss_mib":1151.5390625},{"t_s":23,"rss_mib":1151.5390625},{"t_s":23,"rss_mib":1151.54296875},{"t_s":23,"rss_mib":1151.54296875},{"t_s":23,"rss_mib":1151.54296875},{"t_s":24,"rss_mib":1151.55078125},{"t_s":24,"rss_mib":1151.55078125},{"t_s":24,"rss_mib":1151.56640625},{"t_s":24,"rss_mib":1151.58203125},{"t_s":25,"rss_mib":1151.609375},{"t_s":25,"rss_mib":1151.609375},{"t_s":26,"rss_mib":1151.609375},{"t_s":26,"rss_mib":1151.62109375},{"t_s":26,"rss_mib":1151.62109375},{"t_s":26,"rss_mib":1151.62890625},{"t_s":27,"rss_mib":1151.63671875},{"t_s":27,"rss_mib":1151.63671875},{"t_s":27,"rss_mib":1151.63671875},{"t_s":28,"rss_mib":1151.64453125},{"t_s":28,"rss_mib":1151.64453125},{"t_s":29,"rss_mib":1151.64453125},{"t_s":29,"rss_mib":1151.65234375},{"t_s":29,"rss_mib":1151.65234375},{"t_s":30,"rss_mib":1151.65234375},{"t_s":30,"rss_mib":1151.66015625},{"t_s":30,"rss_mib":1151.66015625},{"t_s":30,"rss_mib":1151.66796875},{"t_s":30,"rss_mib":1151.66796875},{"t_s":31,"rss_mib":1151.66796875},{"t_s":31,"rss_mib":1151.67578125},{"t_s":32,"rss_mib":1151.67578125},{"t_s":32,"rss_mib":1151.68359375},{"t_s":32,"rss_mib":1151.68359375},{"t_s":33,"rss_mib":1151.68359375},{"t_s":33,"rss_mib":1151.703125},{"t_s":33,"rss_mib":1151.703125},{"t_s":34,"rss_mib":1151.703125},{"t_s":35,"rss_mib":1151.703125},{"t_s":35,"rss_mib":1151.73046875},{"t_s":35,"rss_mib":1151.73828125},{"t_s":36,"rss_mib":1151.74609375},{"t_s":36,"rss_mib":1151.74609375},{"t_s":36,"rss_mib":1151.75390625},{"t_s":37,"rss_mib":1151.76171875},{"t_s":37,"rss_mib":1151.76171875},{"t_s":37,"rss_mib":1151.76171875},{"t_s":38,"rss_mib":1151.765625},{"t_s":38,"rss_mib":1151.765625},{"t_s":38,"rss_mib":1151.7734375},{"t_s":39,"rss_mib":1151.7734375},{"t_s":39,"rss_mib":1151.78125},{"t_s":39,"rss_mib":1151.78125},{"t_s":39,"rss_mib":1151.7890625},{"t_s":40,"rss_mib":1151.7890625},{"t_s":40,"rss_mib":1151.79296875},{"t_s":40,"rss_mib":1151.79296875},{"t_s":40,"rss_mib":1151.80078125},{"t_s":40,"rss_mib":1151.80078125},{"t_s":41,"rss_mib":1151.80859375},{"t_s":41,"rss_mib":1151.80859375},{"t_s":42,"rss_mib":1151.80859375},{"t_s":42,"rss_mib":1151.80859375},{"t_s":43,"rss_mib":1151.80859375},{"t_s":43,"rss_mib":1151.81640625},{"t_s":43,"rss_mib":1151.81640625},{"t_s":43,"rss_mib":1151.82421875},{"t_s":43,"rss_mib":1151.83203125},{"t_s":44,"rss_mib":1151.83203125},{"t_s":45,"rss_mib":1151.83984375},{"t_s":45,"rss_mib":1151.83984375},{"t_s":46,"rss_mib":1151.84765625},{"t_s":46,"rss_mib":1151.86328125},{"t_s":46,"rss_mib":1151.86328125},{"t_s":46,"rss_mib":1151.87109375},{"t_s":46,"rss_mib":1151.87109375},{"t_s":47,"rss_mib":1151.87109375},{"t_s":47,"rss_mib":1151.87890625},{"t_s":48,"rss_mib":1151.87890625},{"t_s":48,"rss_mib":1151.8828125},{"t_s":48,"rss_mib":1151.8828125},{"t_s":48,"rss_mib":1151.890625},{"t_s":49,"rss_mib":1151.890625},{"t_s":49,"rss_mib":1151.8984375},{"t_s":49,"rss_mib":1151.8984375},{"t_s":49,"rss_mib":1151.90625},{"t_s":49,"rss_mib":1151.90625},{"t_s":50,"rss_mib":1152.06640625},{"t_s":50,"rss_mib":1152.06640625},{"t_s":50,"rss_mib":1152.078125},{"t_s":50,"rss_mib":1152.078125},{"t_s":51,"rss_mib":1152.08203125},{"t_s":51,"rss_mib":1152.10546875},{"t_s":52,"rss_mib":1152.10546875},{"t_s":52,"rss_mib":1152.26171875},{"t_s":52,"rss_mib":1152.2734375},{"t_s":53,"rss_mib":1152.2734375},{"t_s":53,"rss_mib":1152.28125},{"t_s":53,"rss_mib":1152.28125},{"t_s":54,"rss_mib":1152.2890625},{"t_s":54,"rss_mib":1152.296875},{"t_s":55,"rss_mib":1152.296875},{"t_s":55,"rss_mib":1152.296875},{"t_s":56,"rss_mib":1152.3125},{"t_s":56,"rss_mib":1152.3203125},{"t_s":56,"rss_mib":1152.3203125},{"t_s":56,"rss_mib":1152.3203125},{"t_s":57,"rss_mib":1152.33203125},{"t_s":57,"rss_mib":1152.33203125},{"t_s":58,"rss_mib":1152.33203125},{"t_s":58,"rss_mib":1152.33984375},{"t_s":58,"rss_mib":1152.34765625},{"t_s":59,"rss_mib":1152.34765625},{"t_s":59,"rss_mib":1152.3515625},{"t_s":59,"rss_mib":1152.3515625},{"t_s":59,"rss_mib":1152.359375},{"t_s":59,"rss_mib":1152.359375}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.010346788,"cost":6.036309395,"memory":435.910419446,"streaming":2.725827256,"streams_sustained":686.926559004,"throughput":338.256118339},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.shape":{"reason":"not_measured","detail":"a settled window has no unsettled shape to describe"},"perf.frontier.1ms.concurrency":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.p99_us":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.rps":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"perf.frontier.1ms.samples":{"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"}}},"openai-responses":{"served":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","status":"","path":"/v1/responses","verdict_note":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}}}}},"model":null,"upstream_endpoint":null,"ootb_config":null,"arch":"arm64","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","rig":{"arch":"arm64","release_url":"file:///home/ubuntu/benchmarking/rig","engine":{"commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","dirty":false},"mock":{"origin":"release","sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","asset_updated_at":null},"ugen":null,"box_qualify":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"somaxconn":65535,"tcp_max_syn_backlog":65535},"measured_at":"2026-11-10T17:39:42Z","started_at":null,"finished_at":null,"duration_s":null,"phase_s":null,"build_env":null},"matrix_from_snapshot":true,"snapshot_file":"result_higress_2026-11-10T17-39-42Z.json","ootb_config":"# data/ingresses/anthropic.gen.yaml\napiVersion: networking.k8s.io/v1\nkind: Ingress\nmetadata:\n  name: anthropic\n  namespace: higress-system\n  annotations:\n    higress.io/destination: mock.static\nspec:\n  rules:\n  - host: anthropic.bench\n    http:\n      paths:\n      - path: /\n        pathType: Prefix\n        backend:\n          resource:\n            apiGroup: networking.higress.io\n            kind: McpBridge\n            name: default\n\n\n# data/ingresses/openai-responses.gen.yaml\napiVersion: networking.k8s.io/v1\nkind: Ingress\nmetadata:\n  name: openai-responses\n  namespace: higress-system\n  annotations:\n    higress.io/destination: mock.static\nspec:\n  rules:\n  - host: openai-responses.bench\n    http:\n      paths:\n      - path: /\n        pathType: Prefix\n        backend:\n          resource:\n            apiGroup: networking.higress.io\n            kind: McpBridge\n            name: default\n\n\n# data/ingresses/bedrock.gen.yaml\napiVersion: networking.k8s.io/v1\nkind: Ingress\nmetadata:\n  name: bedrock\n  namespace: higress-system\n  annotations:\n    higress.io/destination: mock.static\nspec:\n  rules:\n  - host: bedrock.bench\n    http:\n      paths:\n      - path: /\n        pathType: Prefix\n        backend:\n          resource:\n            apiGroup: networking.higress.io\n            kind: McpBridge\n            name: default\n\n\n# data/ingresses/gemini.gen.yaml\napiVersion: networking.k8s.io/v1\nkind: Ingress\nmetadata:\n  name: gemini\n  namespace: higress-system\n  annotations:\n    higress.io/destination: mock.static\nspec:\n  rules:\n  - host: gemini.bench\n    http:\n      paths:\n      - path: /\n        pathType: Prefix\n        backend:\n          resource:\n            apiGroup: networking.higress.io\n            kind: McpBridge\n            name: default\n\n\n# data/ingresses/openai.gen.yaml\napiVersion: networking.k8s.io/v1\nkind: Ingress\nmetadata:\n  name: openai\n  namespace: higress-system\n  annotations:\n    higress.io/destination: mock.static\nspec:\n  rules:\n  - http:\n      paths:\n      - path: /\n        pathType: Prefix\n        backend:\n          resource:\n            apiGroup: networking.higress.io\n            kind: McpBridge\n            name: default\n\n\n# data/ingresses/cohere.gen.yaml\napiVersion: networking.k8s.io/v1\nkind: Ingress\nmetadata:\n  name: cohere\n  namespace: higress-system\n  annotations:\n    higress.io/destination: mock.static\nspec:\n  rules:\n  - host: cohere.bench\n    http:\n      paths:\n      - path: /\n        pathType: Prefix\n        backend:\n          resource:\n            apiGroup: networking.higress.io\n            kind: McpBridge\n            name: default\n\n\n# data/wasmplugins/bench.gen.yaml\napiVersion: extensions.higress.io/v1alpha1\nkind: WasmPlugin\nmetadata:\n  name: bench\n  namespace: higress-system\nspec:\n  defaultConfigDisable: true\n  matchRules:\n  - ingress: [openai]\n    config:\n      provider:\n        type: openai\n        apiTokens: [dummy]\n        modelMapping:\n          \"*\": gpt-4o-mini\n  - ingress: [openai-responses]\n    config:\n      provider:\n        type: openai\n        apiTokens: [dummy]\n        modelMapping:\n          \"*\": gpt-4o-mini\n  - ingress: [anthropic]\n    config:\n      provider:\n        type: claude\n        apiTokens: [dummy]\n        modelMapping:\n          \"*\": claude-3-5-sonnet-20241022\n  - ingress: [gemini]\n    config:\n      provider:\n        type: gemini\n        apiTokens: [dummy]\n        modelMapping:\n          \"*\": gemini-1.5-pro\n  - ingress: [cohere]\n    config:\n      provider:\n        type: cohere\n        apiTokens: [dummy]\n        modelMapping:\n          \"*\": command-r-plus\n  - ingress: [bedrock]\n    config:\n      provider:\n        type: bedrock\n        apiTokens: [dummy]\n        modelMapping:\n          \"*\": anthropic.claude-3-sonnet-20240229-v1:0\n        awsRegion: us-east-1\n  url: file:///usr/share/nginx/html/plugins/ai-proxy/2.0.3/plugin.wasm\n\n\n# data/mcpbridges/default.gen.yaml\n# The upstream registry this run drives: the harness's mock, on the host loopback. This file\n# replaces the image's own default McpBridge (whose only registry is the console at 8001); the\n# all-in-one bootstrap copies its defaults into /data with cp -n, so a file that already exists at\n# this path is left alone and this one wins. 127.0.0.1 reaches the mock because the harness\n# launches every container with host networking - the same fact apisix's override.endpoint and\n# kong's upstream_url rely on. Shape mirrors the image's own all-in-one/config/mcpbridges/default.yaml.\napiVersion: networking.higress.io/v1\nkind: McpBridge\nmetadata:\n  name: default\n  namespace: higress-system\n  creationTimestamp: \"2000-01-01T00:00:00Z\"\n  resourceVersion: \"1\"\nspec:\n  registries:\n  - domain: \"127.0.0.1:8000\"\n    name: mock\n    port: 8000\n    type: static\n","best_cell":{"path":{"ingress":"openai","egress":"openai","dialect":"openai"},"source":{"kind":"matrix","sweep":"6x6-diagonal","build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","measured_at":"2026-11-10T17:39:42Z"},"added_latency_p50_us":{"value":718,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":943,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":971,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":709.3196158019531,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1409.801699716714,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.7330563381485045,"certified":true,"suppressed":false},"cost_window_ok":{"value":24883,"certified":true,"suppressed":false},"cost_window_rps":{"value":4146,"certified":true,"suppressed":false},"cost_threads":{"value":142,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.008399308765020295,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":2688,"certified":true,"suppressed":false},"concurrency":2,"p99_us":944,"first_disqualified_conc":4,"lower_bound":false,"samples":16131},{"bound_ms":5,"rps":{"value":5685,"certified":true,"suppressed":false},"concurrency":8,"p99_us":2498,"first_disqualified_conc":16,"lower_bound":false,"samples":34120},{"bound_ms":10,"rps":{"value":6095,"certified":true,"suppressed":false},"concurrency":32,"p99_us":8960,"first_disqualified_conc":64,"lower_bound":false,"samples":36576},{"bound_ms":50,"rps":{"value":6138,"certified":true,"suppressed":false},"concurrency":64,"p99_us":14447,"first_disqualified_conc":256,"lower_bound":false,"samples":36832},{"bound_ms":100,"rps":{"value":6138,"certified":true,"suppressed":false},"concurrency":64,"p99_us":14447,"first_disqualified_conc":512,"lower_bound":false,"samples":36832},{"bound_ms":null,"rps":{"value":6138,"certified":true,"suppressed":false},"concurrency":64,"p99_us":14447,"first_disqualified_conc":null,"lower_bound":false,"samples":36832}],"sweep":[{"conc":1,"ok":8277,"rps":1379,"p99_us":936,"fail":0},{"conc":1,"ok":8251,"rps":1374,"p99_us":942,"fail":0},{"conc":1,"ok":8197,"rps":1365,"p99_us":953,"fail":0},{"conc":2,"ok":15569,"rps":2594,"p99_us":966,"fail":0},{"conc":2,"ok":15751,"rps":2624,"p99_us":957,"fail":0},{"conc":2,"ok":16131,"rps":2688,"p99_us":944,"fail":0},{"conc":4,"ok":25079,"rps":4179,"p99_us":1611,"fail":0},{"conc":4,"ok":24950,"rps":4157,"p99_us":1591,"fail":0},{"conc":4,"ok":31773,"rps":5294,"p99_us":1068,"fail":0},{"conc":8,"ok":26516,"rps":4418,"p99_us":2587,"fail":0},{"conc":8,"ok":26041,"rps":4339,"p99_us":3958,"fail":0},{"conc":8,"ok":34120,"rps":5685,"p99_us":2498,"fail":0},{"conc":16,"ok":35427,"rps":5903,"p99_us":6656,"fail":0},{"conc":16,"ok":34776,"rps":5795,"p99_us":6635,"fail":0},{"conc":16,"ok":34142,"rps":5689,"p99_us":7386,"fail":0},{"conc":32,"ok":36220,"rps":6036,"p99_us":10493,"fail":0},{"conc":32,"ok":36576,"rps":6095,"p99_us":8960,"fail":0},{"conc":32,"ok":36312,"rps":6051,"p99_us":10425,"fail":0},{"conc":64,"ok":36832,"rps":6138,"p99_us":14447,"fail":0},{"conc":64,"ok":36784,"rps":6130,"p99_us":14497,"fail":0},{"conc":64,"ok":36765,"rps":6126,"p99_us":14741,"fail":0},{"conc":128,"ok":36609,"rps":6100,"p99_us":33886,"fail":0},{"conc":128,"ok":36625,"rps":6102,"p99_us":33878,"fail":0},{"conc":128,"ok":36583,"rps":6095,"p99_us":28695,"fail":0},{"conc":256,"ok":34863,"rps":5809,"p99_us":60903,"fail":0},{"conc":256,"ok":35402,"rps":5899,"p99_us":65344,"fail":0},{"conc":256,"ok":35191,"rps":5864,"p99_us":57589,"fail":0},{"conc":512,"ok":33836,"rps":5638,"p99_us":118651,"fail":0},{"conc":512,"ok":33751,"rps":5624,"p99_us":110609,"fail":0},{"conc":512,"ok":33893,"rps":5648,"p99_us":118801,"fail":0},{"conc":1024,"ok":33606,"rps":5599,"p99_us":1034532,"fail":0},{"conc":1024,"ok":33944,"rps":5656,"p99_us":227956,"fail":0},{"conc":1024,"ok":33912,"rps":5650,"p99_us":249044,"fail":0},{"conc":2048,"ok":33804,"rps":5632,"p99_us":810707,"fail":0},{"conc":2048,"ok":33833,"rps":5637,"p99_us":749523,"fail":0},{"conc":2048,"ok":33997,"rps":5664,"p99_us":855634,"fail":0},{"conc":4096,"ok":33921,"rps":5651,"p99_us":772468,"fail":0},{"conc":4096,"ok":33888,"rps":5647,"p99_us":406404,"fail":0},{"conc":4096,"ok":33990,"rps":5663,"p99_us":708027,"fail":0},{"conc":8192,"ok":34130,"rps":5688,"p99_us":653786,"fail":0},{"conc":8192,"ok":34049,"rps":5673,"p99_us":597661,"fail":0},{"conc":8192,"ok":34215,"rps":5701,"p99_us":809671,"fail":0},{"conc":16384,"ok":34112,"rps":5684,"p99_us":677993,"fail":0},{"conc":16384,"ok":34143,"rps":5690,"p99_us":664790,"fail":0},{"conc":16384,"ok":34041,"rps":5671,"p99_us":761220,"fail":0},{"conc":32768,"ok":26116,"rps":4352,"p99_us":3063275,"fail":0},{"conc":32768,"ok":30806,"rps":5133,"p99_us":2651630,"fail":0},{"conc":32768,"ok":29598,"rps":4932,"p99_us":2602395,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1374,"certified":true,"suppressed":false},"p99_us":942,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":2624,"certified":true,"suppressed":false},"p99_us":957,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":4179,"certified":true,"suppressed":false},"p99_us":1591,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":4418,"certified":true,"suppressed":false},"p99_us":2587,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":5795,"certified":true,"suppressed":false},"p99_us":6656,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":6051,"certified":true,"suppressed":false},"p99_us":10425,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":6130,"certified":true,"suppressed":false},"p99_us":14497,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":6100,"certified":true,"suppressed":false},"p99_us":33878,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":5864,"certified":true,"suppressed":false},"p99_us":60903,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":5638,"certified":true,"suppressed":false},"p99_us":118651,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":5650,"certified":true,"suppressed":false},"p99_us":249044,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":5637,"certified":true,"suppressed":false},"p99_us":810707,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":5651,"certified":true,"suppressed":false},"p99_us":708027,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":5688,"certified":true,"suppressed":false},"p99_us":653786,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":5684,"certified":true,"suppressed":false},"p99_us":677993,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":4932,"certified":true,"suppressed":false},"p99_us":2651630,"windows":3,"clean_windows":3}],"reverify_note":"openai>openai is a same-dialect cell: there is no translation to prove, so a request-shape check upstream cannot distinguish a translating gateway from a forwarding one","c1_note":"the c=1 percentiles are taken over 7996 successful gateway round trip(s) and 262593 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":7996,"direct_c1_samples":262593,"rps_per_dollar":{"value":37346.813725490196,"certified":true,"suppressed":false},"cost_per_million_usd":{"value":0.007437790538692918,"certified":true,"suppressed":false},"priced_at_bound_ms":10,"gateway_hourly_usd":0.1632},"translation_cell":{"path":{"ingress":"openai","egress":"anthropic"},"source":{"kind":"matrix","sweep":"6x6-translation","build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","measured_at":"2026-11-10T17:39:42Z"},"added_latency_p50_us":{"value":859,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":1112,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":1140,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":882.4409213281484,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":1133.2203389830509,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.980004881341195,"certified":true,"suppressed":false},"cost_window_ok":{"value":26744,"certified":true,"suppressed":false},"cost_window_rps":{"value":4456,"certified":true,"suppressed":false},"cost_threads":{"value":145,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0.001121746933891714,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":4788,"certified":true,"suppressed":false},"concurrency":16,"p99_us":4880,"first_disqualified_conc":32,"lower_bound":false,"samples":28736},{"bound_ms":10,"rps":{"value":4909,"certified":true,"suppressed":false},"concurrency":32,"p99_us":9079,"first_disqualified_conc":64,"lower_bound":false,"samples":29464},{"bound_ms":50,"rps":{"value":4944,"certified":true,"suppressed":false},"concurrency":64,"p99_us":19849,"first_disqualified_conc":256,"lower_bound":false,"samples":29673},{"bound_ms":100,"rps":{"value":4944,"certified":true,"suppressed":false},"concurrency":64,"p99_us":19849,"first_disqualified_conc":512,"lower_bound":false,"samples":29673},{"bound_ms":null,"rps":{"value":4944,"certified":true,"suppressed":false},"concurrency":64,"p99_us":19849,"first_disqualified_conc":null,"lower_bound":false,"samples":29673}],"sweep":[{"conc":1,"ok":6662,"rps":1110,"p99_us":1165,"fail":0},{"conc":1,"ok":6671,"rps":1111,"p99_us":1166,"fail":0},{"conc":1,"ok":6785,"rps":1130,"p99_us":1133,"fail":0},{"conc":2,"ok":12938,"rps":2155,"p99_us":1171,"fail":0},{"conc":2,"ok":12705,"rps":2117,"p99_us":1176,"fail":0},{"conc":2,"ok":12841,"rps":2139,"p99_us":1176,"fail":0},{"conc":4,"ok":13747,"rps":2290,"p99_us":3082,"fail":0},{"conc":4,"ok":20477,"rps":3412,"p99_us":1981,"fail":0},{"conc":4,"ok":26332,"rps":4387,"p99_us":1225,"fail":0},{"conc":8,"ok":27475,"rps":4578,"p99_us":4112,"fail":0},{"conc":8,"ok":27950,"rps":4657,"p99_us":2973,"fail":0},{"conc":8,"ok":27927,"rps":4653,"p99_us":3024,"fail":0},{"conc":16,"ok":28736,"rps":4788,"p99_us":4880,"fail":0},{"conc":16,"ok":28806,"rps":4800,"p99_us":5046,"fail":0},{"conc":16,"ok":28816,"rps":4802,"p99_us":5836,"fail":0},{"conc":32,"ok":29464,"rps":4909,"p99_us":9079,"fail":0},{"conc":32,"ok":29366,"rps":4894,"p99_us":11175,"fail":0},{"conc":32,"ok":29443,"rps":4906,"p99_us":10943,"fail":0},{"conc":64,"ok":29673,"rps":4944,"p99_us":19849,"fail":0},{"conc":64,"ok":29649,"rps":4939,"p99_us":18297,"fail":0},{"conc":64,"ok":29487,"rps":4914,"p99_us":18823,"fail":0},{"conc":128,"ok":29380,"rps":4895,"p99_us":37219,"fail":0},{"conc":128,"ok":29457,"rps":4908,"p99_us":35728,"fail":0},{"conc":128,"ok":29448,"rps":4907,"p99_us":34166,"fail":0},{"conc":256,"ok":27758,"rps":4625,"p99_us":73977,"fail":0},{"conc":256,"ok":27774,"rps":4627,"p99_us":71715,"fail":0},{"conc":256,"ok":27781,"rps":4629,"p99_us":75422,"fail":0},{"conc":512,"ok":27235,"rps":4538,"p99_us":147749,"fail":0},{"conc":512,"ok":27237,"rps":4539,"p99_us":151100,"fail":0},{"conc":512,"ok":27085,"rps":4514,"p99_us":145713,"fail":0},{"conc":1024,"ok":26932,"rps":4487,"p99_us":1877868,"fail":0},{"conc":1024,"ok":27228,"rps":4537,"p99_us":497956,"fail":0},{"conc":1024,"ok":27403,"rps":4566,"p99_us":1259616,"fail":0},{"conc":2048,"ok":26743,"rps":4456,"p99_us":1080307,"fail":0},{"conc":2048,"ok":26981,"rps":4496,"p99_us":1118256,"fail":0},{"conc":2048,"ok":26969,"rps":4494,"p99_us":902839,"fail":0},{"conc":4096,"ok":26965,"rps":4493,"p99_us":1029012,"fail":0},{"conc":4096,"ok":27017,"rps":4501,"p99_us":1007242,"fail":0},{"conc":4096,"ok":26705,"rps":4449,"p99_us":816316,"fail":0},{"conc":8192,"ok":26756,"rps":4459,"p99_us":890982,"fail":0},{"conc":8192,"ok":26783,"rps":4463,"p99_us":1048045,"fail":0},{"conc":8192,"ok":27035,"rps":4504,"p99_us":1013440,"fail":0},{"conc":16384,"ok":27078,"rps":4512,"p99_us":834839,"fail":0},{"conc":16384,"ok":27256,"rps":4542,"p99_us":863669,"fail":0},{"conc":16384,"ok":26890,"rps":4480,"p99_us":758820,"fail":0},{"conc":32768,"ok":21394,"rps":3565,"p99_us":2922297,"fail":0},{"conc":32768,"ok":23790,"rps":3964,"p99_us":2570015,"fail":0},{"conc":32768,"ok":19677,"rps":3279,"p99_us":2759758,"fail":0}],"rungs":[{"conc":1,"rps":{"value":1111,"certified":true,"suppressed":false},"p99_us":1165,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":2139,"certified":true,"suppressed":false},"p99_us":1176,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":3412,"certified":true,"suppressed":false},"p99_us":1981,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":4653,"certified":true,"suppressed":false},"p99_us":3024,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":4800,"certified":true,"suppressed":false},"p99_us":5046,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":4906,"certified":true,"suppressed":false},"p99_us":10943,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":4939,"certified":true,"suppressed":false},"p99_us":18823,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":4907,"certified":true,"suppressed":false},"p99_us":35728,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":4627,"certified":true,"suppressed":false},"p99_us":73977,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":4538,"certified":true,"suppressed":false},"p99_us":147749,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":4537,"certified":true,"suppressed":false},"p99_us":1259616,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":4494,"certified":true,"suppressed":false},"p99_us":1080307,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":4493,"certified":true,"suppressed":false},"p99_us":1007242,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":4463,"certified":true,"suppressed":false},"p99_us":1013440,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":4512,"certified":true,"suppressed":false},"p99_us":834839,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":3565,"certified":true,"suppressed":false},"p99_us":2759758,"windows":3,"clean_windows":3}],"egress_reverified":true,"c1_note":"the c=1 percentiles are taken over 6739 successful gateway round trip(s) and 262868 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":6739,"direct_c1_samples":262868},"streaming":{"path":{"dialect":"openai"},"source":{"kind":"matrix","sweep":"6x6-stream-diagonal","build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","measured_at":"2026-11-10T17:39:42Z"},"stream_served":true,"added_ttft_p50_us":{"value":881,"certified":true,"suppressed":false},"added_ttft_p99_us":{"value":1103,"certified":true,"suppressed":false},"added_gap_p50_us":{"value":2,"certified":true,"suppressed":false},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"below_resolution","detail":"the gateway's own inter-frame gap at this percentile (21141us) came in under the mock's (21153us), which a proxy cannot really do - the added gap here is below what this rig can resolve against the mock's 20ms pacing"},"streams_sustained_fps":{"value":52911.767627714544,"certified":true,"suppressed":false,"headroom":0.617121104958904,"rig_ceiling":85739.68253968254},"streams_sustained":{"value":1688,"certified":true,"suppressed":false,"headroom":0.617121104958904,"rig_ceiling":85739.68253968254},"ttft_gw_samples":100,"ttft_direct_samples":100,"gap_gw_samples":63,"gap_direct_samples":63},"measured_at":"2026-11-10T17:39:42.000Z","lane_measured_at":{"perf":"2026-11-10T17:39:42Z","xlate":"2026-11-10T17:39:42Z","stream":"2026-11-10T17:39:42Z","memory":"2026-11-10T17:39:42Z"},"rig":{"arch":"arm64","release_url":"file:///home/ubuntu/benchmarking/rig","engine":{"commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","dirty":false},"mock":{"origin":"release","sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","asset_updated_at":null},"ugen":null,"box_qualify":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"somaxconn":65535,"tcp_max_syn_backlog":65535},"contributed_by":[{"handle":"geekspeng","name":"geekspeng","url":"https://github.com/geekspeng"}],"stale":false,"engine":{"sha":"d240badd966933fa19eec0606b2d1f3945f18ad2","short":"d240bad","current":true},"history_family":"higress","history_trends":{}},"current_metadata":{"stars":9764,"stars_as_of":"2026-11-14","first_commit":"2022-10-27"},"metadata_scope":"gateway.stars and stars_as_of are legacy aliases of current_metadata, a dated repository snapshot, not stars at measurement time. Run statistics were not recorded.","measurements":[{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:cohere:cohere","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"cohere","upstream_protocol":"cohere","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":615,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":747,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":774,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":27,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":613.2194813409235,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1630.737493553378,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.8053044884180178,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":31620,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":5269,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":143,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.0076217583807716635,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer cohere with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1207.94921875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1217.2890625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1218.421875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1226.8046875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1324.41015625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":0.7482121208635604,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":1.0318152244070045,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":3092,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":2,"p99_us":790,"first_disqualified_conc":4,"lower_bound":false,"samples":18555},{"bound_ms":5,"rps":{"value":5921,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":8,"p99_us":2939,"first_disqualified_conc":16,"lower_bound":false,"samples":35536},{"bound_ms":10,"rps":{"value":6529,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":32,"p99_us":9093,"first_disqualified_conc":64,"lower_bound":false,"samples":39189},{"bound_ms":50,"rps":{"value":6633,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":18236,"first_disqualified_conc":256,"lower_bound":false,"samples":39805},{"bound_ms":100,"rps":{"value":6633,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":18236,"first_disqualified_conc":512,"lower_bound":false,"samples":39805},{"bound_ms":null,"rps":{"value":6633,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":18236,"first_disqualified_conc":null,"lower_bound":false,"samples":39805}],"sample_counts":{"gateway_c1_samples":9341,"direct_c1_samples":267984},"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:bedrock:cohere","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"bedrock","upstream_protocol":"cohere","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:gemini:cohere","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"gemini","upstream_protocol":"cohere","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai:cohere","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_served","raw_state":"failed","reason":"failed","detail":"probed and answered HTTP 200, but the body is not a usable openai response (see body_snippet)","evidence_scope":"Recorded run evidence","probe_http_status":"200","response_excerpt":"the 2xx body is not a response of the openai dialect with assistant text at choices[].message.content: {\"response_id\":\"x\",\"generation_id\":\"x\",\"text\":\"ok\",\"finish_reason\":\"COMPLETE\",\"meta\":{\"api_version\":{\"version\":\"1\"},\"tokens\":{\"input_tokens\":10,\"output_tokens\":2},\"billed_units\":{\"input_tokens\":10,\"output_tokens\":2}}}"},"test":{"ingress_protocol":"openai","upstream_protocol":"cohere","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:anthropic:cohere","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_served","raw_state":"failed","reason":"failed","detail":"probed and answered HTTP 200, but the body is not a usable anthropic response (see body_snippet)","evidence_scope":"Recorded run evidence","probe_http_status":"200","response_excerpt":"the 2xx body is not a response of the anthropic dialect with assistant text at content[type=text].text: {\"id\":\"\",\"type\":\"message\",\"role\":\"assistant\",\"content\":null,\"model\":\"\",\"stop_reason\":null,\"stop_sequence\":null,\"usage\":{},\"error\":null}"},"test":{"ingress_protocol":"anthropic","upstream_protocol":"cohere","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai-responses:cohere","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai-responses","upstream_protocol":"cohere","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:cohere:anthropic","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"cohere","upstream_protocol":"anthropic","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:gemini:anthropic","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"gemini","upstream_protocol":"anthropic","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai:anthropic","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai","upstream_protocol":"anthropic","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":859,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":1112,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":1140,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":882.4409213281484,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1133.2203389830509,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.980004881341195,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":26744,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":4456,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":145,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.001121746933891714,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":1095,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":1556,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"below_resolution","detail":"the gateway's own inter-frame gap at this percentile (21173us) came in under the mock's (21174us), which a proxy cannot really do - the added gap here is below what this rig can resolve against the mock's 20ms pacing","state":"below_resolution","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":23656.559777267295,"certified":true,"suppressed":false,"headroom":0.08075923714495402,"rig_ceiling":292926.98412698414,"state":"valid","reason":null,"section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":5767,"certified":true,"suppressed":false,"headroom":0.08075923714495402,"rig_ceiling":292926.98412698414,"state":"valid","reason":null,"section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1148.421875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1158.875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1159.7734375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1167.30859375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1264.890625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":0.7021128067897227,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":0.8890053062912313,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound","state":"bound_not_met"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":4788,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":4880,"first_disqualified_conc":32,"lower_bound":false,"samples":28736},{"bound_ms":10,"rps":{"value":4909,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":32,"p99_us":9079,"first_disqualified_conc":64,"lower_bound":false,"samples":29464},{"bound_ms":50,"rps":{"value":4944,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":19849,"first_disqualified_conc":256,"lower_bound":false,"samples":29673},{"bound_ms":100,"rps":{"value":4944,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":19849,"first_disqualified_conc":512,"lower_bound":false,"samples":29673},{"bound_ms":null,"rps":{"value":4944,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":19849,"first_disqualified_conc":null,"lower_bound":false,"samples":29673}],"sample_counts":{"gateway_c1_samples":6739,"direct_c1_samples":262868,"ttft_gw_samples":100,"ttft_direct_samples":100,"gap_gw_samples":63,"gap_direct_samples":63},"stream_rung_diagnostics":[{"conc":1,"http_status_counts":{"200":1},"stream_errors_no_response":0},{"conc":2,"http_status_counts":{"200":2},"stream_errors_no_response":0},{"conc":4,"http_status_counts":{"200":4},"stream_errors_no_response":0},{"conc":8,"http_status_counts":{"200":8},"stream_errors_no_response":0},{"conc":16,"http_status_counts":{"200":16},"stream_errors_no_response":0},{"conc":32,"http_status_counts":{"200":32},"stream_errors_no_response":0},{"conc":64,"http_status_counts":{"200":64},"stream_errors_no_response":0},{"conc":128,"http_status_counts":{"200":128},"stream_errors_no_response":0},{"conc":256,"http_status_counts":{"200":256},"stream_errors_no_response":0},{"conc":512,"http_status_counts":{"200":512},"stream_errors_no_response":0},{"conc":1024,"http_status_counts":{"200":1024},"stream_errors_no_response":0},{"conc":2048,"http_status_counts":{"200":2048},"stream_errors_no_response":0},{"conc":4096,"http_status_counts":{"200":4096},"stream_errors_no_response":0},{"conc":8192,"http_status_counts":{"200":8082},"stream_errors_no_response":110},{"conc":6144,"http_status_counts":{"200":6144},"stream_errors_no_response":0},{"conc":7168,"http_status_counts":{"200":7168},"stream_errors_no_response":0},{"conc":7680,"http_status_counts":{"200":7680},"stream_errors_no_response":0},{"conc":7424,"http_status_counts":{"200":7424},"stream_errors_no_response":0},{"conc":7552,"http_status_counts":{"200":7552},"stream_errors_no_response":0},{"conc":7488,"http_status_counts":{"200":7488},"stream_errors_no_response":0},{"conc":7456,"http_status_counts":{"200":7456},"stream_errors_no_response":0},{"conc":7440,"http_status_counts":{"200":7440},"stream_errors_no_response":0},{"conc":7432,"http_status_counts":{"200":7432},"stream_errors_no_response":0},{"conc":7436,"http_status_counts":{"200":7436},"stream_errors_no_response":0},{"conc":7438,"http_status_counts":{"200":7438},"stream_errors_no_response":0},{"conc":7439,"http_status_counts":{"200":7439},"stream_errors_no_response":0},{"conc":7438,"http_status_counts":{"200":7438},"stream_errors_no_response":0},{"conc":7438,"http_status_counts":{"200":7438},"stream_errors_no_response":0},{"conc":5767,"http_status_counts":{"200":5767},"stream_errors_no_response":0},{"conc":5767,"http_status_counts":{"200":5767},"stream_errors_no_response":0},{"conc":5767,"http_status_counts":{"200":5767},"stream_errors_no_response":0}],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:bedrock:anthropic","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"bedrock","upstream_protocol":"anthropic","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai-responses:anthropic","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai-responses","upstream_protocol":"anthropic","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:anthropic:anthropic","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"anthropic","upstream_protocol":"anthropic","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":680,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":822,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":850,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":674.8307904095445,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1481.852953676158,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.9771911881943663,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":34868,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":5810,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":144,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.0020075714121830907,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":822,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":1012,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":35,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":52509.65359935977,"certified":true,"suppressed":false,"headroom":0.06307020958070865,"rig_ceiling":832558.7301587302,"state":"valid","reason":null,"section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":16391,"certified":true,"suppressed":false,"headroom":0.06307020958070865,"rig_ceiling":832558.7301587302,"state":"valid","reason":null,"section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1153.921875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1161.83984375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1162.21484375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1171.8125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1272.52734375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":0.275204255009843,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":0.5817656189913647,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":1352,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":1,"p99_us":957,"first_disqualified_conc":2,"lower_bound":false,"samples":8115},{"bound_ms":5,"rps":{"value":5967,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":4103,"first_disqualified_conc":32,"lower_bound":false,"samples":35809},{"bound_ms":10,"rps":{"value":6123,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":32,"p99_us":9073,"first_disqualified_conc":64,"lower_bound":false,"samples":36743},{"bound_ms":50,"rps":{"value":6224,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":128,"p99_us":27064,"first_disqualified_conc":256,"lower_bound":false,"samples":37352},{"bound_ms":100,"rps":{"value":6224,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":128,"p99_us":27064,"first_disqualified_conc":512,"lower_bound":false,"samples":37352},{"bound_ms":null,"rps":{"value":6224,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":128,"p99_us":27064,"first_disqualified_conc":null,"lower_bound":false,"samples":37352}],"sample_counts":{"gateway_c1_samples":8481,"direct_c1_samples":262245,"ttft_gw_samples":100,"ttft_direct_samples":100,"gap_gw_samples":63,"gap_direct_samples":63},"stream_rung_diagnostics":[{"conc":1,"http_status_counts":{"200":1},"stream_errors_no_response":0},{"conc":2,"http_status_counts":{"200":2},"stream_errors_no_response":0},{"conc":4,"http_status_counts":{"200":4},"stream_errors_no_response":0},{"conc":8,"http_status_counts":{"200":8},"stream_errors_no_response":0},{"conc":16,"http_status_counts":{"200":16},"stream_errors_no_response":0},{"conc":32,"http_status_counts":{"200":32},"stream_errors_no_response":0},{"conc":64,"http_status_counts":{"200":64},"stream_errors_no_response":0},{"conc":128,"http_status_counts":{"200":128},"stream_errors_no_response":0},{"conc":256,"http_status_counts":{"200":256},"stream_errors_no_response":0},{"conc":512,"http_status_counts":{"200":512},"stream_errors_no_response":0},{"conc":1024,"http_status_counts":{"200":1024},"stream_errors_no_response":0},{"conc":2048,"http_status_counts":{"200":2048},"stream_errors_no_response":0},{"conc":4096,"http_status_counts":{"200":4096},"stream_errors_no_response":0},{"conc":8192,"http_status_counts":{"200":8192},"stream_errors_no_response":0},{"conc":16384,"http_status_counts":{"200":16384},"stream_errors_no_response":0},{"conc":32768,"http_status_counts":{"200":17771},"stream_errors_no_response":14997},{"conc":24576,"http_status_counts":{"200":17562},"stream_errors_no_response":7014},{"conc":20480,"http_status_counts":{"200":17305},"stream_errors_no_response":3175},{"conc":18432,"http_status_counts":{"200":17810},"stream_errors_no_response":622},{"conc":17408,"http_status_counts":{"200":17408},"stream_errors_no_response":0},{"conc":16896,"http_status_counts":{"200":16896},"stream_errors_no_response":0},{"conc":16640,"http_status_counts":{"200":16640},"stream_errors_no_response":0},{"conc":16512,"http_status_counts":{"200":16512},"stream_errors_no_response":0},{"conc":16448,"http_status_counts":{"200":16448},"stream_errors_no_response":0},{"conc":16416,"http_status_counts":{"200":16416},"stream_errors_no_response":0},{"conc":16400,"http_status_counts":{"200":16400},"stream_errors_no_response":0},{"conc":16392,"http_status_counts":{"200":16392},"stream_errors_no_response":0},{"conc":16388,"http_status_counts":{"200":16388},"stream_errors_no_response":0},{"conc":16390,"http_status_counts":{"200":16390},"stream_errors_no_response":0},{"conc":16391,"http_status_counts":{"200":16391},"stream_errors_no_response":0},{"conc":16391,"http_status_counts":{"200":16391},"stream_errors_no_response":0},{"conc":16391,"http_status_counts":{"200":16391},"stream_errors_no_response":0}],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai:openai-responses","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai","upstream_protocol":"openai-responses","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:gemini:openai-responses","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"gemini","upstream_protocol":"openai-responses","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:cohere:openai-responses","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"cohere","upstream_protocol":"openai-responses","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:bedrock:openai-responses","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"bedrock","upstream_protocol":"openai-responses","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai-responses:openai-responses","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai-responses","upstream_protocol":"openai-responses","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":642,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":747,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":775,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":633.7934164508637,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1577.8011794439765,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.9859729286272246,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":37457,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":6242,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":141,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.0010144966227941374,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer openai-responses with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1147.16796875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1156.55078125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1157.92578125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1165.46875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1186.5546875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":66,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":0.5987504753473016,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":0.6971452873324683,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":2925,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":2,"p99_us":890,"first_disqualified_conc":4,"lower_bound":false,"samples":17553},{"bound_ms":5,"rps":{"value":6438,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":3521,"first_disqualified_conc":32,"lower_bound":false,"samples":38636},{"bound_ms":10,"rps":{"value":6717,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":32,"p99_us":6634,"first_disqualified_conc":64,"lower_bound":false,"samples":40311},{"bound_ms":50,"rps":{"value":6791,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":128,"p99_us":26658,"first_disqualified_conc":256,"lower_bound":false,"samples":40754},{"bound_ms":100,"rps":{"value":6791,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":128,"p99_us":26658,"first_disqualified_conc":512,"lower_bound":false,"samples":40754},{"bound_ms":null,"rps":{"value":6791,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":128,"p99_us":26658,"first_disqualified_conc":null,"lower_bound":false,"samples":40754}],"sample_counts":{"gateway_c1_samples":8971,"direct_c1_samples":260207},"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:anthropic:openai-responses","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"anthropic","upstream_protocol":"openai-responses","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai:gemini","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai","upstream_protocol":"gemini","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":923,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":1226,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":1254,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":951.6544794570992,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1050.8015478164732,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.7511344229974055,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":19009,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":3167,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":141,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.011047398600662844,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1150.5859375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1160.25390625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1160.609375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1170.90234375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1268.890625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":0.25904735694337694,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":0.6054154815784856,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound","state":"bound_not_met"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":4171,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":8,"p99_us":4447,"first_disqualified_conc":16,"lower_bound":false,"samples":25035},{"bound_ms":10,"rps":{"value":4327,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":6571,"first_disqualified_conc":32,"lower_bound":false,"samples":25969},{"bound_ms":50,"rps":{"value":4449,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":22764,"first_disqualified_conc":256,"lower_bound":false,"samples":26699},{"bound_ms":100,"rps":{"value":4449,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":22764,"first_disqualified_conc":512,"lower_bound":false,"samples":26699},{"bound_ms":null,"rps":{"value":4449,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":22764,"first_disqualified_conc":null,"lower_bound":false,"samples":26699}],"sample_counts":{"gateway_c1_samples":6239,"direct_c1_samples":265561},"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:gemini:gemini","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"gemini","upstream_protocol":"gemini","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":750,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":932,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":959,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":27,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":737.715063745513,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1355.5369127516778,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.7426132259918599,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":24237,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":4038,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":140,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.007014069398027809,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1156.921875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1165.796875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1166.64453125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1175.75390625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1269.25,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":66,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":1.08419810667365,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":0.8945988670581346,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound","state":"bound_not_met"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":5241,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":4564,"first_disqualified_conc":32,"lower_bound":false,"samples":31452},{"bound_ms":10,"rps":{"value":5428,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":32,"p99_us":8261,"first_disqualified_conc":64,"lower_bound":false,"samples":32577},{"bound_ms":50,"rps":{"value":5506,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":16888,"first_disqualified_conc":256,"lower_bound":false,"samples":33045},{"bound_ms":100,"rps":{"value":5506,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":16888,"first_disqualified_conc":512,"lower_bound":false,"samples":33045},{"bound_ms":null,"rps":{"value":5506,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":16888,"first_disqualified_conc":null,"lower_bound":false,"samples":33045}],"sample_counts":{"gateway_c1_samples":7714,"direct_c1_samples":265544},"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:bedrock:gemini","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"bedrock","upstream_protocol":"gemini","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:cohere:gemini","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"cohere","upstream_protocol":"gemini","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai-responses:gemini","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai-responses","upstream_protocol":"gemini","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:anthropic:gemini","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"anthropic","upstream_protocol":"gemini","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":1045,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":1360,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":1388,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":1055.3296800424253,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":947.571189279732,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.7438065009330738,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":16971,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":2828,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":142,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.01019386011431265,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer gemini with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1151.53125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1164.61328125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1165.828125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1173.3046875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1280.23828125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":1.1668804149685401,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":1.20687057093733,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound","state":"bound_not_met"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":3821,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":8,"p99_us":2813,"first_disqualified_conc":16,"lower_bound":false,"samples":22929},{"bound_ms":10,"rps":{"value":3862,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":8825,"first_disqualified_conc":32,"lower_bound":false,"samples":23185},{"bound_ms":50,"rps":{"value":3961,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":26997,"first_disqualified_conc":256,"lower_bound":false,"samples":23771},{"bound_ms":100,"rps":{"value":3961,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":26997,"first_disqualified_conc":512,"lower_bound":false,"samples":23771},{"bound_ms":null,"rps":{"value":3961,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":26997,"first_disqualified_conc":null,"lower_bound":false,"samples":23771}],"sample_counts":{"gateway_c1_samples":5488,"direct_c1_samples":262925},"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai-responses:bedrock","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai-responses","upstream_protocol":"bedrock","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:gemini:bedrock","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"gemini","upstream_protocol":"bedrock","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:cohere:bedrock","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"cohere","upstream_protocol":"bedrock","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:bedrock:bedrock","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"bedrock","upstream_protocol":"bedrock","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":727,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":861,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":889,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":778.1517645882627,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1285.0963597430407,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.9697792765122447,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":30007,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":5000,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":143,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.005065484720231946,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1157.3984375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1165.7421875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1167.375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1175.8359375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1275.18359375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":0.6896154469149554,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":0.6918641708843657,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound","state":"bound_not_met"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":4996,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":8,"p99_us":3109,"first_disqualified_conc":16,"lower_bound":false,"samples":29985},{"bound_ms":10,"rps":{"value":5179,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":6789,"first_disqualified_conc":32,"lower_bound":false,"samples":31083},{"bound_ms":50,"rps":{"value":5403,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":21391,"first_disqualified_conc":256,"lower_bound":false,"samples":32432},{"bound_ms":100,"rps":{"value":5403,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":21391,"first_disqualified_conc":512,"lower_bound":false,"samples":32432},{"bound_ms":null,"rps":{"value":5403,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":21391,"first_disqualified_conc":null,"lower_bound":false,"samples":32432}],"sample_counts":{"gateway_c1_samples":7954,"direct_c1_samples":263620},"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai:bedrock","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai","upstream_protocol":"bedrock","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":929,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":1235,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":1263,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":912.3461771232601,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1096.0751796572692,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.7513336009437701,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":19828,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":3304,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":144,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.008069396812588258,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":null,"certified":false,"suppressed":false,"reason":"untestable","detail":"the mock does not answer bedrock with a native event stream, so the rig cannot pose the streaming question here","state":"untestable","section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1158.6875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1172.09765625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1173.9765625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1179.87109375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1198.6875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":0.9309596103361604,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":1.0752406083257144,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound","state":"bound_not_met"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":4316,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":8,"p99_us":3304,"first_disqualified_conc":16,"lower_bound":false,"samples":25900},{"bound_ms":10,"rps":{"value":4440,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":5213,"first_disqualified_conc":32,"lower_bound":false,"samples":26646},{"bound_ms":50,"rps":{"value":4587,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":20787,"first_disqualified_conc":256,"lower_bound":false,"samples":27530},{"bound_ms":100,"rps":{"value":4587,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":20787,"first_disqualified_conc":512,"lower_bound":false,"samples":27530},{"bound_ms":null,"rps":{"value":4587,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":20787,"first_disqualified_conc":null,"lower_bound":false,"samples":27530}],"sample_counts":{"gateway_c1_samples":6162,"direct_c1_samples":262294},"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:anthropic:bedrock","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"anthropic","upstream_protocol":"bedrock","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai:openai","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai","upstream_protocol":"openai","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":718,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":943,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":971,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":709.3196158019531,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1409.801699716714,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.7330563381485045,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":24883,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":4146,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":142,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.008399308765020295,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":881,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":1103,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":2,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"below_resolution","detail":"the gateway's own inter-frame gap at this percentile (21141us) came in under the mock's (21153us), which a proxy cannot really do - the added gap here is below what this rig can resolve against the mock's 20ms pacing","state":"below_resolution","section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":52911.767627714544,"certified":true,"suppressed":false,"headroom":0.617121104958904,"rig_ceiling":85739.68253968254,"state":"valid","reason":null,"section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":1688,"certified":true,"suppressed":false,"headroom":0.617121104958904,"rig_ceiling":85739.68253968254,"state":"valid","reason":null,"section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1163.541015625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1170.94140625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1171.9921875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1181.8125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1287.14453125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":66,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":0.8963186649029332,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":0.6061112080903336,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":2688,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":2,"p99_us":944,"first_disqualified_conc":4,"lower_bound":false,"samples":16131},{"bound_ms":5,"rps":{"value":5685,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":8,"p99_us":2498,"first_disqualified_conc":16,"lower_bound":false,"samples":34120},{"bound_ms":10,"rps":{"value":6095,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":32,"p99_us":8960,"first_disqualified_conc":64,"lower_bound":false,"samples":36576},{"bound_ms":50,"rps":{"value":6138,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":14447,"first_disqualified_conc":256,"lower_bound":false,"samples":36832},{"bound_ms":100,"rps":{"value":6138,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":14447,"first_disqualified_conc":512,"lower_bound":false,"samples":36832},{"bound_ms":null,"rps":{"value":6138,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":14447,"first_disqualified_conc":null,"lower_bound":false,"samples":36832}],"sample_counts":{"gateway_c1_samples":7996,"direct_c1_samples":262593,"ttft_gw_samples":100,"ttft_direct_samples":100,"gap_gw_samples":63,"gap_direct_samples":63},"stream_rung_diagnostics":[{"conc":1,"http_status_counts":{"200":1},"stream_errors_no_response":0},{"conc":2,"http_status_counts":{"200":2},"stream_errors_no_response":0},{"conc":4,"http_status_counts":{"200":4},"stream_errors_no_response":0},{"conc":8,"http_status_counts":{"200":8},"stream_errors_no_response":0},{"conc":16,"http_status_counts":{"200":16},"stream_errors_no_response":0},{"conc":32,"http_status_counts":{"200":32},"stream_errors_no_response":0},{"conc":64,"http_status_counts":{"200":64},"stream_errors_no_response":0},{"conc":128,"http_status_counts":{"200":128},"stream_errors_no_response":0},{"conc":256,"http_status_counts":{"200":256},"stream_errors_no_response":0},{"conc":512,"http_status_counts":{"200":512},"stream_errors_no_response":0},{"conc":1024,"http_status_counts":{"200":1024},"stream_errors_no_response":0},{"conc":2048,"http_status_counts":{"200":2048},"stream_errors_no_response":0},{"conc":1536,"http_status_counts":{"200":1536},"stream_errors_no_response":0},{"conc":1792,"http_status_counts":{"200":1792},"stream_errors_no_response":0},{"conc":1664,"http_status_counts":{"200":1664},"stream_errors_no_response":0},{"conc":1728,"http_status_counts":{"200":1728},"stream_errors_no_response":0},{"conc":1696,"http_status_counts":{"200":1696},"stream_errors_no_response":0},{"conc":1680,"http_status_counts":{"200":1680},"stream_errors_no_response":0},{"conc":1688,"http_status_counts":{"200":1688},"stream_errors_no_response":0},{"conc":1692,"http_status_counts":{"200":1692},"stream_errors_no_response":0},{"conc":1690,"http_status_counts":{"200":1690},"stream_errors_no_response":0},{"conc":1689,"http_status_counts":{"200":1689},"stream_errors_no_response":0},{"conc":1688,"http_status_counts":{"200":1688},"stream_errors_no_response":0},{"conc":1688,"http_status_counts":{"200":1688},"stream_errors_no_response":0}],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:gemini:openai","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"gemini","upstream_protocol":"openai","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:bedrock:openai","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"bedrock","upstream_protocol":"openai","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:cohere:openai","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"cohere","upstream_protocol":"openai","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:anthropic:openai","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"served","raw_state":true,"reason":null,"detail":null,"evidence_scope":"Recorded run evidence","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"anthropic","upstream_protocol":"openai","upstream_transport":"http","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":"cold restart, 60s idle read at rest, then load at c=32 for 300s, then 60s with the load removed. `plateaued` is a verdict over the trailing 60s, taken at the END of the load: steady means drift under 1% and range under 2% of the mean. The verdict does not end the load - every cell is loaded for the full duration.","memory_concurrency":null,"memory_load_s":300,"cost_concurrency":8,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{"added_latency_p50_us":{"value":954,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"added_latency_p99_us":{"value":1252,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"gateway_c1_p99_us":{"value":1280,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole."},"cpu_us_per_request":{"value":968.311271560369,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"rps_per_cpu_second":{"value":1032.7257663628832,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_conc":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_core_utilisation":{"value":0.5011834686494262,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_ok":{"value":12465,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_window_rps":{"value":2076,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_threads":{"value":142,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_nonvol_ctxt_per_request":{"value":0.00008022462896109105,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"cost_majflt":{"value":0,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"perf","definition":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else."},"added_ttft_p50_us":{"value":1262,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_ttft_p99_us":{"value":1705,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p50_us":{"value":8,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"added_gap_p99_us":{"value":163,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"stream","definition":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added."},"streams_sustained_fps":{"value":19376.729819513093,"certified":true,"suppressed":false,"headroom":0.07497629094372328,"rig_ceiling":258438.09523809524,"state":"valid","reason":null,"section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"streams_sustained":{"value":5088,"certified":true,"suppressed":false,"headroom":0.07497629094372328,"rig_ceiling":258438.09523809524,"state":"valid","reason":null,"section":"stream","definition":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"idle_rss_mib":{"value":1151.65234375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"steady_state_rss_mib":{"value":1160.24609375,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"recovered_rss_mib":{"value":1161.1953125,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_mib":{"value":1177.22265625,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"peak_rss_hwm_mib":{"value":1280.44921875,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"time_to_plateau_s":{"value":60,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"growth_rate_mib_per_min":{"value":0.5290055438736946,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"load_s":{"value":300,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_shape":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe","state":"not_measured","section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_static":{"value":1,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."},"idle_growth_rate_mib_per_min":{"value":1.003040818815334,"certified":true,"suppressed":false,"state":"valid","reason":null,"section":"memory","definition":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran."}},"frontier":[{"bound_ms":1,"rps":{"value":null,"certified":false,"suppressed":false,"reason":"bound_not_met","detail":"every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound","state":"bound_not_met"},"concurrency":null,"p99_us":null,"first_disqualified_conc":null,"lower_bound":false},{"bound_ms":5,"rps":{"value":3867,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":8,"p99_us":4852,"first_disqualified_conc":16,"lower_bound":false,"samples":23211},{"bound_ms":10,"rps":{"value":3999,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":5892,"first_disqualified_conc":32,"lower_bound":false,"samples":23998},{"bound_ms":50,"rps":{"value":4118,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":23767,"first_disqualified_conc":256,"lower_bound":false,"samples":24713},{"bound_ms":100,"rps":{"value":4118,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":23767,"first_disqualified_conc":512,"lower_bound":false,"samples":24713},{"bound_ms":null,"rps":{"value":4118,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":64,"p99_us":23767,"first_disqualified_conc":null,"lower_bound":false,"samples":24713}],"sample_counts":{"gateway_c1_samples":6058,"direct_c1_samples":259722,"ttft_gw_samples":100,"ttft_direct_samples":100,"gap_gw_samples":63,"gap_direct_samples":63},"stream_rung_diagnostics":[{"conc":1,"http_status_counts":{"200":1},"stream_errors_no_response":0},{"conc":2,"http_status_counts":{"200":2},"stream_errors_no_response":0},{"conc":4,"http_status_counts":{"200":4},"stream_errors_no_response":0},{"conc":8,"http_status_counts":{"200":8},"stream_errors_no_response":0},{"conc":16,"http_status_counts":{"200":16},"stream_errors_no_response":0},{"conc":32,"http_status_counts":{"200":32},"stream_errors_no_response":0},{"conc":64,"http_status_counts":{"200":64},"stream_errors_no_response":0},{"conc":128,"http_status_counts":{"200":128},"stream_errors_no_response":0},{"conc":256,"http_status_counts":{"200":256},"stream_errors_no_response":0},{"conc":512,"http_status_counts":{"200":512},"stream_errors_no_response":0},{"conc":1024,"http_status_counts":{"200":1024},"stream_errors_no_response":0},{"conc":2048,"http_status_counts":{"200":2048},"stream_errors_no_response":0},{"conc":4096,"http_status_counts":{"200":4096},"stream_errors_no_response":0},{"conc":8192,"http_status_counts":{"200":6688},"stream_errors_no_response":1504},{"conc":6144,"http_status_counts":{"200":6144},"stream_errors_no_response":0},{"conc":5120,"http_status_counts":{"200":5120},"stream_errors_no_response":0},{"conc":5632,"http_status_counts":{"200":5632},"stream_errors_no_response":0},{"conc":5888,"http_status_counts":{"200":5888},"stream_errors_no_response":0},{"conc":6016,"http_status_counts":{"200":6016},"stream_errors_no_response":0},{"conc":6080,"http_status_counts":{"200":6080},"stream_errors_no_response":0},{"conc":6112,"http_status_counts":{"200":6112},"stream_errors_no_response":0},{"conc":6096,"http_status_counts":{"200":6096},"stream_errors_no_response":0},{"conc":6088,"http_status_counts":{"200":6088},"stream_errors_no_response":0},{"conc":6084,"http_status_counts":{"200":6084},"stream_errors_no_response":0},{"conc":6082,"http_status_counts":{"200":6082},"stream_errors_no_response":0},{"conc":6081,"http_status_counts":{"200":6081},"stream_errors_no_response":0},{"conc":6080,"http_status_counts":{"200":6080},"stream_errors_no_response":0},{"conc":6080,"http_status_counts":{"200":6080},"stream_errors_no_response":0},{"conc":5088,"http_status_counts":{"200":5088},"stream_errors_no_response":0},{"conc":5088,"http_status_counts":{"200":5088},"stream_errors_no_response":0},{"conc":5088,"http_status_counts":{"200":5088},"stream_errors_no_response":0}],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"},{"schema_version":"1.0","measurement_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270:openai-responses:openai","gateway_id":"higress","gateway_build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","run_id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","measured_at":"2026-11-10T17:39:42Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","detail":"Higress 2.2.5 with bundled AI Proxy 2.0.3. Source 4deae1dd3b4cb2707273748985e55bef28e747d1 and pinned-image probes establish 13 configured exact-wire pairs; two Cohere translations failed response correctness in the preflight. In the 2026-11-09 run the engine withheld performance and streaming numbers for six configured pairs because re-verification found the request did not arrive on the requested upstream wire: OpenAI to Anthropic, Gemini, Cohere and Bedrock arrived untranslated on the mock's OpenAI Chat Completions endpoint, and Anthropic to Gemini and Anthropic to Cohere also arrived on the OpenAI Chat Completions endpoint. Responses cross-conversion is not implemented; Messages-to-Bedrock uses Anthropic Mantle rather than Converse. Pairs outside the configured 13 are not configured: unsupported translations pass through the original wire or return unsupported API errors. Evidence: results/diagnostics/refresh-configuration/higress-adapter-preflight.json","evidence_scope":"Historical capability research, not a fresh probe of the tested build. Referenced versions may differ from gateway_build; do not infer current capability.","probe_http_status":null,"response_excerpt":null},"test":{"ingress_protocol":"openai-responses","upstream_protocol":"openai","hardware":"AWS m7g.4xlarge (Graviton3, 16 cores / 64 GB). Gateway-under-test pinned to 4 cores (the comparable basis); mock and load generator on 6 cores each so the mock never bottlenecks the streaming sweep. Ubuntu 24.04. One dedicated box per gateway.","architecture":"arm64","harness_commit":"d240badd966933fa19eec0606b2d1f3945f18ad2","harness_dirty":false,"mock_sha256":"99d01f709d8d9c1f49055aa265d5eb7475264d6b513911fe317aee74a71190f8","load_generator_sha256":null,"qualification":{"band_pct":20,"baseline_rps":96199,"baseline_samples":1,"concurrency":32,"drift_pct":0.6995914718448216,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":96872,"outcome":"pass"},"definitions":{"memory":"See `matrix.memory.protocol` and each cell's `memory.protocol`, which state the durations and both plateau thresholds formatted from the constants that ran.","perf.added_latency":"WHAT THE GATEWAY ADDS, at concurrency 1. The gateway leg's percentile minus the same request taken straight to the mock, one at a time, both legs in the same window. No search and no threshold: there is nothing to stop early and nothing to decide. A difference at or below what the rig can resolve publishes `below_resolution` - the best result the comparison can express - and every surface renders that apart from a never-measured hole.","perf.cost":"WHAT A REQUEST COST, in microseconds of gateway CPU. User plus system time, summed across the gateway's whole PROCESS TREE (a gateway that forks workers keeps its CPU in the children, and a parent-only reading would report a busy gateway as idle), sampled as the DIFFERENCE across one load window and divided by the requests that window actually completed. WHY IT EXISTS BESIDE THROUGHPUT. Peak rps is a SATURATION number: once a gateway fills the cores it is pinned to, the ladder stops describing the gateway and starts describing the box, and two gateways at that wall read the same however different they are. Cost per request has no such ceiling - at saturation both serve the same rate by definition, and the one doing less work per request still reads lower. It is also the figure that maps to money, since half the CPU serves the same traffic on half the instance. MEASURED AT ONE CONCURRENCY, c=8, HELD IDENTICAL FOR EVERY GATEWAY and published as `cost_window_conc`. No single concurrency is below saturation for a field spanning double-digit to five-figure rps, so matched LOAD is impossible and matched CONCURRENCY is the honest substitute - and it is stated rather than assumed. REFUSED, NOT ESTIMATED, IN TWO CASES. A window with ANY failure publishes no cost at all: CPU spent refusing requests is real, but dividing it by only the successes would describe the failures rather than the work. And a window with major page faults means the box was SWAPPING, so what was timed is the disk - those numbers publish with `cost_majflt` non-zero and the cost itself marked a harness fault, so a reader sees why the row looks wrong instead of finding a hole. `cost_core_utilisation` is the fraction of the PINNED cores busy across that same window. It is what makes a peak interpretable: near 1.0 the gateway had filled the cores it was given and its throughput number is a real ceiling; well below it, the limit is somewhere else and the peak means something else.","perf.frontier":"THROUGHPUT AT A TAIL LATENCY YOU ACCEPT. For each declared bound, the most requests/sec this cell carried while 99% of requests finished under that bound AND it failed none it accepted. Bounds: 1ms, 5ms, 10ms, 50ms, 100ms plus one reading with no latency bound at all (failures only). ONE MEASUREMENT, READ SIX WAYS: a single concurrency sweep, published as `sweep_max_proxy`, so every reading is re-derivable from the rungs rather than taken on trust. Monotone non-decreasing across the bounds BY CONSTRUCTION - relaxing a bound only adds rungs to the set a maximum is taken over - so a reading cannot exceed a looser one. ZERO FAILURES, not a tolerance: the load generator counts connections the RIG could not open separately (`rig_refused`, and those windows are discarded), so a failure reaching a reading is the gateway losing a request it had accepted. TERMINATION: the sweep climbs a doubling ladder from the floor and stops at the first concurrency where no window served cleanly, because more concurrency cannot un-fail those requests. Each reading names `first_disqualified_conc`, the lowest concurrency above it that stopped qualifying - that pair is the proof it is a boundary. When nothing above it disqualified, the sweep ran out of range instead of finding a limit, and the reading sets `lower_bound: true`: the rate is real, it is simply a floor rather than a ceiling. A window in which the gateway completed no request at all is a failed rung like any other. Only a RIG stop (a thread or connection this host could not make, or a load generator with no readable result) interrupts the sweep; its frontier is then read from the rungs below the stop, every reading sets `lower_bound: true`, and the absences name the rung and the cause. A bound no rung met publishes `bound_not_met`. The `p99_us` on a reading is the tail the winning rung ACTUALLY produced, never the bound - 4ms under a 100ms bound is not the same finding as 99ms - and `samples` is the number of completed requests it was taken over.","stream.added_ttft_and_gap":"WHAT THE GATEWAY ADDS TO A STREAM, at concurrency 1: time to the first event, and the gaps between frames inside one stream, each as the gateway leg minus the same stream taken straight to the mock. Time-to-first-EVENT rather than first token, deliberately and by the same definition on both legs: a frame budget of one is satisfied by a dialect's opening scaffolding (openai sends a role delta, anthropic a `message_start`), so the quantity is the same on both sides and the difference still isolates what the gateway added.","stream.streams_sustained":"THE MOST CONCURRENT SSE STREAMS THIS CELL CARRIES CLEANLY, and the frames/sec at that point. A concurrency holds the gate when every expected content frame arrives (ratio 1.0, i.e. no tolerance - a proxy that drops a frame has dropped a user's token), no inter-frame gap exceeds 10x the mock's own pacing interval, and under 0.1% of streams error. TERMINATION: the gate is monotone in concurrency, so the ceiling is a boundary found by bisection - a pass proven at n and a failure measured above it. `sweep_streams` carries every rung probed. A range whose top still passed publishes an absence rather than a ceiling, because the top of the range is our choice and not the gateway's answer."},"sweep_duration_s":null,"sweep_ttft_ms":null,"memory_protocol":null,"memory_concurrency":null,"memory_load_s":null,"cost_concurrency":null,"somaxconn":65535,"tcp_max_syn_backlog":65535},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/higress","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270"},"disclosure":{"operator":"Busbar team","operator_is_entrant":true},"published_at":"2026-11-14T11:09:23.855Z","generated_at":"2026-11-14T11:09:24.858Z","age_days":3,"age_as_of":"2026-11-14T11:09:24.858Z","age":"3 days old at collection generation"}],"history":[{"id":"run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","sha256":"558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","gateway_id":"higress","measured_at":"2026-11-10T17:39:42Z","build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422","gateway_family":"higress","engine":"d240badd966933fa19eec0606b2d1f3945f18ad2","projector":"historical-matrix-v1-e1440c114a2bba3bde4a01bfa138bc6160e327576b1d93bd2dced418d56805ae","configuration_sha256":"74a072c4dcf00a0a6363d16a46c32b39846d5c07fae45ebe06272191267ccfe5","disposition":"recorded","evidence_available":true,"available_metrics":["added_gap_p50_us","added_gap_p99_us","added_latency_p50_us","added_latency_p99_us","added_ttft_p50_us","added_ttft_p99_us","cost_core_utilisation","cost_majflt","cost_nonvol_ctxt_per_request","cost_threads","cost_window_conc","cost_window_ok","cost_window_rps","cpu_us_per_request","direct_c1_p99_us","gateway_c1_p99_us","growth_rate_mib_per_min","idle_growth_rate_mib_per_min","idle_rss_mib","idle_shape","idle_static","load_s","peak_rss_hwm_mib","peak_rss_mib","recovered_rss_mib","rps_per_cpu_second","shape","steady_state_rss_mib","streams_sustained","streams_sustained_fps","time_to_plateau_s"],"measurements_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/measurements","source_url":"https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270/raw.json"},{"id":"run-higress-af1ce2bee0833f93ed558ceb54c4e30f0164de90a7a837817aa387b275cc353b","sha256":"af1ce2bee0833f93ed558ceb54c4e30f0164de90a7a837817aa387b275cc353b","gateway_id":"higress","measured_at":"2026-08-26T19:01:18Z","build":"higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.4","gateway_family":"higress","engine":"f3adbb1315b26129f5e317af5279decefb1cea8f","projector":"historical-matrix-v1-e1440c114a2bba3bde4a01bfa138bc6160e327576b1d93bd2dced418d56805ae","configuration_sha256":"12f3ba4d67834925fb4f01dca07a798610f552aac41d28fc1d4798213134ebd7","disposition":"recorded","evidence_available":true,"available_metrics":["added_gap_p50_us","added_gap_p99_us","added_latency_p50_us","added_latency_p99_us","added_ttft_p50_us","added_ttft_p99_us","cost_core_utilisation","cost_majflt","cost_nonvol_ctxt_per_request","cost_threads","cost_window_conc","cost_window_ok","cost_window_rps","cpu_us_per_request","direct_c1_p99_us","gateway_c1_p99_us","growth_rate_mib_per_min","idle_growth_rate_mib_per_min","idle_rss_mib","idle_shape","idle_static","load_s","peak_rss_hwm_mib","peak_rss_mib","recovered_rss_mib","rps_per_cpu_second","shape","steady_state_rss_mib","streams_sustained","streams_sustained_fps","time_to_plateau_s"],"measurements_url":"https://onthebench.ai/api/v1/runs/run-higress-af1ce2bee0833f93ed558ceb54c4e30f0164de90a7a837817aa387b275cc353b/measurements","source_url":"https://onthebench.ai/gateways/runs/run-higress-af1ce2bee0833f93ed558ceb54c4e30f0164de90a7a837817aa387b275cc353b","raw_result_url":"https://onthebench.ai/api/v1/runs/run-higress-af1ce2bee0833f93ed558ceb54c4e30f0164de90a7a837817aa387b275cc353b/raw.json"}],"generated_at":"2026-11-14T11:09:24.858Z","disclosure":{"operator":"Busbar team","operator_is_entrant":true}}
