{"schema_version":"1.0","gateway":{"key":"litellm-rust","display":"LiteLLM · Rust","lang":"Rust","cls":"LLM gateway","repo":"https://github.com/BerriAI/litellm","version":"v1.108.3","stars":63241,"stars_as_of":"2026-11-14","first_commit":"2023-07-27","matrix":{"gateway":"litellm-rust","build":"src/litellm-rust/target/release/litellm-gateway","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":{"bedrock":{"configurable":true,"served":false,"egress_config":null,"serve_error":"","cells":{"anthropic":{"served":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/messages","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v2/chat","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/model/bench-bedrock/converse","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1beta/models/bench-bedrock:generateContent","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"501","path":"/v1/chat/completions","verdict_note":"probed and answered failed (HTTP 501)","body_snippet":"{\"error\":{\"message\":\"unsupported by the rust path: unrecognized request parameter\",\"type\":\"api_error\",\"param\":null,\"code\":501}}","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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/responses","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":{"bedrock":{"served":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/model/bench-openai/converse","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":249,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":279,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":306,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":27,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":272.6013913681787,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":3668.359853121175,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.6790579018170796,"certified":true,"suppressed":false},"cost_window_ok":{"value":59941,"certified":true,"suppressed":false},"cost_window_rps":{"value":9989,"certified":true,"suppressed":false},"cost_threads":{"value":5,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":8404,"certified":true,"suppressed":false},"concurrency":4,"p99_us":588,"first_disqualified_conc":8,"lower_bound":false,"samples":50435},{"bound_ms":5,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":32,"lower_bound":false,"samples":60306},{"bound_ms":10,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":64,"lower_bound":false,"samples":60306},{"bound_ms":50,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":256,"lower_bound":false,"samples":60306},{"bound_ms":100,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":1024,"lower_bound":false,"samples":60306},{"bound_ms":null,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":null,"lower_bound":false,"samples":60306}],"sweep":[{"conc":1,"ok":22499,"rps":3748,"p99_us":301,"fail":0},{"conc":1,"ok":22443,"rps":3738,"p99_us":300,"fail":0},{"conc":1,"ok":22367,"rps":3726,"p99_us":302,"fail":0},{"conc":2,"ok":32753,"rps":5458,"p99_us":410,"fail":0},{"conc":2,"ok":32736,"rps":5455,"p99_us":409,"fail":0},{"conc":2,"ok":32631,"rps":5437,"p99_us":409,"fail":0},{"conc":4,"ok":50172,"rps":8361,"p99_us":589,"fail":0},{"conc":4,"ok":50353,"rps":8391,"p99_us":593,"fail":0},{"conc":4,"ok":50435,"rps":8404,"p99_us":588,"fail":0},{"conc":8,"ok":60135,"rps":10021,"p99_us":1218,"fail":0},{"conc":8,"ok":60158,"rps":10025,"p99_us":1223,"fail":0},{"conc":8,"ok":60188,"rps":10030,"p99_us":1221,"fail":0},{"conc":16,"ok":60306,"rps":10049,"p99_us":2486,"fail":0},{"conc":16,"ok":60202,"rps":10032,"p99_us":2503,"fail":0},{"conc":16,"ok":60205,"rps":10033,"p99_us":2499,"fail":0},{"conc":32,"ok":59745,"rps":9956,"p99_us":5251,"fail":0},{"conc":32,"ok":59963,"rps":9992,"p99_us":5176,"fail":0},{"conc":32,"ok":59839,"rps":9972,"p99_us":5236,"fail":0},{"conc":64,"ok":59403,"rps":9900,"p99_us":11251,"fail":0},{"conc":64,"ok":59179,"rps":9862,"p99_us":11128,"fail":0},{"conc":64,"ok":59226,"rps":9870,"p99_us":11116,"fail":0},{"conc":128,"ok":59428,"rps":9903,"p99_us":25558,"fail":0},{"conc":128,"ok":59277,"rps":9878,"p99_us":25080,"fail":0},{"conc":128,"ok":59326,"rps":9886,"p99_us":25057,"fail":0},{"conc":256,"ok":59141,"rps":9855,"p99_us":54342,"fail":0},{"conc":256,"ok":59008,"rps":9833,"p99_us":60650,"fail":0},{"conc":256,"ok":58946,"rps":9823,"p99_us":57205,"fail":0},{"conc":512,"ok":57998,"rps":9664,"p99_us":94929,"fail":0},{"conc":512,"ok":58039,"rps":9672,"p99_us":94698,"fail":0},{"conc":512,"ok":58127,"rps":9687,"p99_us":95092,"fail":0},{"conc":1024,"ok":56975,"rps":9494,"p99_us":146358,"fail":0},{"conc":1024,"ok":57324,"rps":9552,"p99_us":151112,"fail":0},{"conc":1024,"ok":57310,"rps":9549,"p99_us":147598,"fail":0},{"conc":2048,"ok":56728,"rps":9453,"p99_us":412827,"fail":0},{"conc":2048,"ok":56479,"rps":9411,"p99_us":344184,"fail":0},{"conc":2048,"ok":56128,"rps":9353,"p99_us":343443,"fail":0},{"conc":4096,"ok":55274,"rps":9211,"p99_us":824938,"fail":0},{"conc":4096,"ok":56042,"rps":9339,"p99_us":898347,"fail":0},{"conc":4096,"ok":55560,"rps":9259,"p99_us":669703,"fail":0},{"conc":8192,"ok":54530,"rps":9087,"p99_us":1884573,"fail":0},{"conc":8192,"ok":54911,"rps":9151,"p99_us":1996218,"fail":0},{"conc":8192,"ok":54551,"rps":9091,"p99_us":2235636,"fail":0},{"conc":16384,"ok":53775,"rps":8961,"p99_us":3860409,"fail":0},{"conc":16384,"ok":53451,"rps":8908,"p99_us":2392444,"fail":0},{"conc":16384,"ok":52935,"rps":8821,"p99_us":2261573,"fail":0},{"conc":32768,"ok":31085,"rps":5180,"p99_us":5817043,"fail":0},{"conc":32768,"ok":16399,"rps":2730,"p99_us":1118348,"fail":0},{"conc":32768,"ok":38059,"rps":6342,"p99_us":5856115,"fail":0}],"rungs":[{"conc":1,"rps":{"value":3738,"certified":true,"suppressed":false},"p99_us":301,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":5455,"certified":true,"suppressed":false},"p99_us":409,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":8391,"certified":true,"suppressed":false},"p99_us":589,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":10025,"certified":true,"suppressed":false},"p99_us":1221,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":10033,"certified":true,"suppressed":false},"p99_us":2499,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":9972,"certified":true,"suppressed":false},"p99_us":5236,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":9870,"certified":true,"suppressed":false},"p99_us":11128,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":9886,"certified":true,"suppressed":false},"p99_us":25080,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":9833,"certified":true,"suppressed":false},"p99_us":57205,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":9672,"certified":true,"suppressed":false},"p99_us":94929,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":9549,"certified":true,"suppressed":false},"p99_us":147598,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":9411,"certified":true,"suppressed":false},"p99_us":344184,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":9259,"certified":true,"suppressed":false},"p99_us":824938,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":9091,"certified":true,"suppressed":false},"p99_us":1996218,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":8908,"certified":true,"suppressed":false},"p99_us":2392444,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":5180,"certified":true,"suppressed":false},"p99_us":5817043,"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 22021 successful gateway round trip(s) and 267021 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":22021,"direct_c1_samples":267021},"stream":{"stream_served":"not_measured","stream_error":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","reason":"not_measured","stream_c1_note":null,"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"streams_sustained_fps":{"value":0,"certified":true,"suppressed":false,"note":"measured_failure"},"streams_sustained":{"value":0,"certified":true,"suppressed":false,"note":"measured_failure"}},"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":12.73828125,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":18.9921875,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":18.99609375,"certified":true,"suppressed":false},"peak_rss_mib":{"value":19.875,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":19.484375,"certified":true,"suppressed":false},"plateaued":false,"time_to_plateau_s":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"memory was still oscillating with no net trend - it returns to where it was, which is not a leak when the 300s load window closed, so it has no time-to-plateau; its rate is published beside this"},"growth_rate_mib_per_min":{"value":-0.1791634760562065,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":0,"certified":true,"suppressed":false},"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,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":13.12890625},{"t_s":2,"rss_mib":17.92578125},{"t_s":2,"rss_mib":17.9296875},{"t_s":5,"rss_mib":17.9609375},{"t_s":5,"rss_mib":17.96875},{"t_s":8,"rss_mib":18.51953125},{"t_s":9,"rss_mib":18.51953125},{"t_s":11,"rss_mib":18.5859375},{"t_s":11,"rss_mib":18.5859375},{"t_s":14,"rss_mib":18.875},{"t_s":14,"rss_mib":18.875},{"t_s":16,"rss_mib":18.87890625},{"t_s":18,"rss_mib":18.87890625},{"t_s":19,"rss_mib":19.33984375},{"t_s":20,"rss_mib":19.33984375},{"t_s":21,"rss_mib":19.34375},{"t_s":23,"rss_mib":19.34375},{"t_s":24,"rss_mib":19.0703125},{"t_s":27,"rss_mib":19.0703125},{"t_s":29,"rss_mib":19.08203125},{"t_s":29,"rss_mib":19.08203125},{"t_s":30,"rss_mib":18.80078125},{"t_s":32,"rss_mib":18.88671875},{"t_s":33,"rss_mib":18.890625},{"t_s":36,"rss_mib":18.890625},{"t_s":37,"rss_mib":18.9140625},{"t_s":38,"rss_mib":18.9140625},{"t_s":42,"rss_mib":18.9140625},{"t_s":44,"rss_mib":18.9296875},{"t_s":45,"rss_mib":18.9296875},{"t_s":47,"rss_mib":18.9296875},{"t_s":48,"rss_mib":18.96875},{"t_s":51,"rss_mib":18.96875},{"t_s":52,"rss_mib":18.97265625},{"t_s":54,"rss_mib":18.97265625},{"t_s":56,"rss_mib":18.97265625},{"t_s":60,"rss_mib":18.97265625},{"t_s":63,"rss_mib":18.97265625},{"t_s":66,"rss_mib":18.97265625},{"t_s":69,"rss_mib":18.97265625},{"t_s":72,"rss_mib":18.97265625},{"t_s":72,"rss_mib":19.8671875},{"t_s":75,"rss_mib":19.8671875},{"t_s":78,"rss_mib":19.875},{"t_s":78,"rss_mib":19.41796875},{"t_s":81,"rss_mib":19.421875},{"t_s":84,"rss_mib":19.421875},{"t_s":87,"rss_mib":19.421875},{"t_s":90,"rss_mib":19.421875},{"t_s":90,"rss_mib":18.86328125},{"t_s":93,"rss_mib":18.8984375},{"t_s":96,"rss_mib":18.8984375},{"t_s":96,"rss_mib":19.484375},{"t_s":99,"rss_mib":19.484375},{"t_s":102,"rss_mib":19.484375},{"t_s":104,"rss_mib":19.65234375},{"t_s":105,"rss_mib":19.65234375},{"t_s":108,"rss_mib":19.65234375},{"t_s":111,"rss_mib":19.65234375},{"t_s":114,"rss_mib":19.65234375},{"t_s":114,"rss_mib":19.66015625},{"t_s":117,"rss_mib":19.66015625},{"t_s":120,"rss_mib":19.66015625},{"t_s":120,"rss_mib":18.921875},{"t_s":123,"rss_mib":18.921875},{"t_s":126,"rss_mib":18.92578125},{"t_s":129,"rss_mib":18.92578125},{"t_s":132,"rss_mib":18.92578125},{"t_s":134,"rss_mib":18.9296875},{"t_s":135,"rss_mib":18.9296875},{"t_s":138,"rss_mib":18.9296875},{"t_s":141,"rss_mib":18.9296875},{"t_s":144,"rss_mib":18.9296875},{"t_s":147,"rss_mib":18.9296875},{"t_s":150,"rss_mib":18.9296875},{"t_s":150,"rss_mib":19.11328125},{"t_s":153,"rss_mib":19.11328125},{"t_s":155,"rss_mib":19.1171875},{"t_s":156,"rss_mib":19.1171875},{"t_s":156,"rss_mib":19.125},{"t_s":159,"rss_mib":19.125},{"t_s":162,"rss_mib":19.125},{"t_s":165,"rss_mib":19.125},{"t_s":168,"rss_mib":19.125},{"t_s":171,"rss_mib":19.12890625},{"t_s":171,"rss_mib":19.12890625},{"t_s":171,"rss_mib":19.1328125},{"t_s":174,"rss_mib":19.1328125},{"t_s":174,"rss_mib":18.890625},{"t_s":177,"rss_mib":18.890625},{"t_s":179,"rss_mib":18.8984375},{"t_s":180,"rss_mib":18.8984375},{"t_s":180,"rss_mib":18.90234375},{"t_s":183,"rss_mib":18.90234375},{"t_s":185,"rss_mib":18.90625},{"t_s":186,"rss_mib":18.90625},{"t_s":189,"rss_mib":18.90625},{"t_s":191,"rss_mib":18.9375},{"t_s":192,"rss_mib":18.9375},{"t_s":192,"rss_mib":18.96875},{"t_s":195,"rss_mib":18.96875},{"t_s":198,"rss_mib":18.96875},{"t_s":201,"rss_mib":18.96875},{"t_s":204,"rss_mib":18.96875},{"t_s":204,"rss_mib":19.19140625},{"t_s":207,"rss_mib":19.19140625},{"t_s":207,"rss_mib":19.1953125},{"t_s":210,"rss_mib":19.1953125},{"t_s":212,"rss_mib":19.19921875},{"t_s":213,"rss_mib":19.19921875},{"t_s":216,"rss_mib":19.19921875},{"t_s":216,"rss_mib":18.98828125},{"t_s":219,"rss_mib":18.98828125},{"t_s":222,"rss_mib":18.98828125},{"t_s":224,"rss_mib":18.9921875},{"t_s":225,"rss_mib":18.9921875},{"t_s":228,"rss_mib":18.9921875},{"t_s":231,"rss_mib":18.9921875},{"t_s":234,"rss_mib":18.9921875},{"t_s":237,"rss_mib":18.9921875},{"t_s":240,"rss_mib":18.9921875},{"t_s":243,"rss_mib":18.9921875},{"t_s":246,"rss_mib":18.9921875},{"t_s":246,"rss_mib":19.4296875},{"t_s":249,"rss_mib":19.4296875},{"t_s":252,"rss_mib":19.453125},{"t_s":252,"rss_mib":18.98046875},{"t_s":255,"rss_mib":18.984375},{"t_s":258,"rss_mib":18.984375},{"t_s":261,"rss_mib":18.984375},{"t_s":264,"rss_mib":18.984375},{"t_s":267,"rss_mib":18.984375},{"t_s":269,"rss_mib":18.98828125},{"t_s":270,"rss_mib":18.98828125},{"t_s":272,"rss_mib":18.9921875},{"t_s":273,"rss_mib":18.9921875},{"t_s":276,"rss_mib":18.9921875},{"t_s":279,"rss_mib":18.9921875},{"t_s":282,"rss_mib":18.9921875},{"t_s":285,"rss_mib":18.9921875},{"t_s":288,"rss_mib":18.9921875},{"t_s":291,"rss_mib":18.9921875},{"t_s":294,"rss_mib":18.9921875},{"t_s":297,"rss_mib":18.9921875},{"t_s":299,"rss_mib":18.99609375},{"t_s":300,"rss_mib":18.99609375},{"t_s":303,"rss_mib":18.99609375},{"t_s":306,"rss_mib":18.99609375},{"t_s":309,"rss_mib":18.99609375},{"t_s":312,"rss_mib":18.99609375},{"t_s":315,"rss_mib":18.99609375},{"t_s":318,"rss_mib":18.99609375},{"t_s":321,"rss_mib":18.99609375},{"t_s":324,"rss_mib":18.99609375},{"t_s":327,"rss_mib":18.99609375},{"t_s":330,"rss_mib":18.99609375},{"t_s":333,"rss_mib":18.99609375},{"t_s":336,"rss_mib":18.99609375},{"t_s":339,"rss_mib":18.99609375},{"t_s":342,"rss_mib":18.99609375},{"t_s":345,"rss_mib":18.99609375},{"t_s":348,"rss_mib":18.99609375},{"t_s":351,"rss_mib":18.99609375},{"t_s":354,"rss_mib":18.99609375},{"t_s":357,"rss_mib":18.99609375},{"t_s":360,"rss_mib":18.99609375}],"idle_rss_series":[{"t_s":0,"rss_mib":12.73828125},{"t_s":0,"rss_mib":12.73828125},{"t_s":0,"rss_mib":12.73828125},{"t_s":1,"rss_mib":12.73828125},{"t_s":1,"rss_mib":12.73828125},{"t_s":2,"rss_mib":12.73828125},{"t_s":2,"rss_mib":12.73828125},{"t_s":3,"rss_mib":12.73828125},{"t_s":3,"rss_mib":12.73828125},{"t_s":4,"rss_mib":12.73828125},{"t_s":5,"rss_mib":12.73828125},{"t_s":5,"rss_mib":12.73828125},{"t_s":5,"rss_mib":12.73828125},{"t_s":6,"rss_mib":12.73828125},{"t_s":6,"rss_mib":12.73828125},{"t_s":7,"rss_mib":12.73828125},{"t_s":7,"rss_mib":12.73828125},{"t_s":8,"rss_mib":12.73828125},{"t_s":8,"rss_mib":12.73828125},{"t_s":9,"rss_mib":12.73828125},{"t_s":10,"rss_mib":12.73828125},{"t_s":10,"rss_mib":12.73828125},{"t_s":10,"rss_mib":12.73828125},{"t_s":11,"rss_mib":12.73828125},{"t_s":11,"rss_mib":12.73828125},{"t_s":12,"rss_mib":12.73828125},{"t_s":12,"rss_mib":12.73828125},{"t_s":13,"rss_mib":12.73828125},{"t_s":13,"rss_mib":12.73828125},{"t_s":14,"rss_mib":12.73828125},{"t_s":15,"rss_mib":12.73828125},{"t_s":15,"rss_mib":12.73828125},{"t_s":16,"rss_mib":12.73828125},{"t_s":16,"rss_mib":12.73828125},{"t_s":16,"rss_mib":12.73828125},{"t_s":17,"rss_mib":12.73828125},{"t_s":17,"rss_mib":12.73828125},{"t_s":18,"rss_mib":12.73828125},{"t_s":18,"rss_mib":12.73828125},{"t_s":19,"rss_mib":12.73828125},{"t_s":20,"rss_mib":12.73828125},{"t_s":20,"rss_mib":12.73828125},{"t_s":21,"rss_mib":12.73828125},{"t_s":21,"rss_mib":12.73828125},{"t_s":21,"rss_mib":12.73828125},{"t_s":22,"rss_mib":12.73828125},{"t_s":22,"rss_mib":12.73828125},{"t_s":23,"rss_mib":12.73828125},{"t_s":23,"rss_mib":12.73828125},{"t_s":24,"rss_mib":12.73828125},{"t_s":25,"rss_mib":12.73828125},{"t_s":25,"rss_mib":12.73828125},{"t_s":26,"rss_mib":12.73828125},{"t_s":26,"rss_mib":12.73828125},{"t_s":27,"rss_mib":12.73828125},{"t_s":27,"rss_mib":12.73828125},{"t_s":27,"rss_mib":12.73828125},{"t_s":28,"rss_mib":12.73828125},{"t_s":28,"rss_mib":12.73828125},{"t_s":29,"rss_mib":12.73828125},{"t_s":30,"rss_mib":12.73828125},{"t_s":30,"rss_mib":12.73828125},{"t_s":31,"rss_mib":12.73828125},{"t_s":31,"rss_mib":12.73828125},{"t_s":32,"rss_mib":12.73828125},{"t_s":32,"rss_mib":12.73828125},{"t_s":32,"rss_mib":12.73828125},{"t_s":33,"rss_mib":12.73828125},{"t_s":33,"rss_mib":12.73828125},{"t_s":34,"rss_mib":12.73828125},{"t_s":35,"rss_mib":12.73828125},{"t_s":35,"rss_mib":12.73828125},{"t_s":36,"rss_mib":12.73828125},{"t_s":36,"rss_mib":12.73828125},{"t_s":37,"rss_mib":12.73828125},{"t_s":37,"rss_mib":12.73828125},{"t_s":38,"rss_mib":12.73828125},{"t_s":38,"rss_mib":12.73828125},{"t_s":38,"rss_mib":12.73828125},{"t_s":39,"rss_mib":12.73828125},{"t_s":39,"rss_mib":12.73828125},{"t_s":40,"rss_mib":12.73828125},{"t_s":41,"rss_mib":12.73828125},{"t_s":41,"rss_mib":12.73828125},{"t_s":42,"rss_mib":12.73828125},{"t_s":42,"rss_mib":12.73828125},{"t_s":43,"rss_mib":12.73828125},{"t_s":43,"rss_mib":12.73828125},{"t_s":43,"rss_mib":12.73828125},{"t_s":44,"rss_mib":12.73828125},{"t_s":44,"rss_mib":12.73828125},{"t_s":45,"rss_mib":12.73828125},{"t_s":46,"rss_mib":12.73828125},{"t_s":46,"rss_mib":12.73828125},{"t_s":47,"rss_mib":12.73828125},{"t_s":47,"rss_mib":12.73828125},{"t_s":48,"rss_mib":12.73828125},{"t_s":48,"rss_mib":12.73828125},{"t_s":49,"rss_mib":12.73828125},{"t_s":49,"rss_mib":12.73828125},{"t_s":49,"rss_mib":12.73828125},{"t_s":50,"rss_mib":12.73828125},{"t_s":51,"rss_mib":12.73828125},{"t_s":51,"rss_mib":12.73828125},{"t_s":52,"rss_mib":12.73828125},{"t_s":52,"rss_mib":12.73828125},{"t_s":53,"rss_mib":12.73828125},{"t_s":53,"rss_mib":12.73828125},{"t_s":54,"rss_mib":12.73828125},{"t_s":54,"rss_mib":12.73828125},{"t_s":54,"rss_mib":12.73828125},{"t_s":55,"rss_mib":12.73828125},{"t_s":56,"rss_mib":12.73828125},{"t_s":56,"rss_mib":12.73828125},{"t_s":57,"rss_mib":12.73828125},{"t_s":57,"rss_mib":12.73828125},{"t_s":58,"rss_mib":12.73828125},{"t_s":58,"rss_mib":12.73828125},{"t_s":59,"rss_mib":12.73828125},{"t_s":59,"rss_mib":12.73828125},{"t_s":59,"rss_mib":12.73828125}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.009090262,"cost":6.02665077,"memory":422.725558526,"streaming":1.306106893,"streams_sustained":0.002167838,"throughput":606.626222704},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.time_to_plateau_s":{"reason":"not_measured","detail":"memory was still oscillating with no net trend - it returns to where it was, which is not a leak when the 300s load window closed, so it has no time-to-plateau; its rate is published beside this"},"stream.added_gap_p50_us":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.added_gap_p99_us":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.added_ttft_p50_us":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.added_ttft_p99_us":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.gap_direct_samples":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.gap_gw_samples":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.streams_sustained_headroom":{"reason":"not_measured","detail":"the sustained figure was measured but the rig could not derive the mock's frame ceiling for this cell, so there is nothing to state this as a fraction of"},"stream.streams_sustained_mock_ceiling":{"reason":"not_measured","detail":"the sustained figure was measured but the rig could not derive the mock's frame ceiling for this cell, so there is nothing to state this as a fraction of"},"stream.ttft_direct_samples":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.ttft_gw_samples":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"}}},"anthropic":{"served":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/messages","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/responses","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1beta/models/bench-openai:generateContent","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v2/chat","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":{"configurable":true,"served":true,"egress_config":null,"serve_error":"","cells":{"gemini":{"served":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1beta/models/bench-responses:generateContent","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v2/chat","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/model/bench-responses/converse","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":267,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":298,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":325,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":27,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":287.74399582885263,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":3475.3114382785957,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.7339219155366118,"certified":true,"suppressed":false},"cost_window_ok":{"value":61374,"certified":true,"suppressed":false},"cost_window_rps":{"value":10227,"certified":true,"suppressed":false},"cost_threads":{"value":5,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":8050,"certified":true,"suppressed":false},"concurrency":4,"p99_us":705,"first_disqualified_conc":8,"lower_bound":false,"samples":48309},{"bound_ms":5,"rps":{"value":10191,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2455,"first_disqualified_conc":32,"lower_bound":false,"samples":61157},{"bound_ms":10,"rps":{"value":10191,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2455,"first_disqualified_conc":64,"lower_bound":false,"samples":61157},{"bound_ms":50,"rps":{"value":10191,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2455,"first_disqualified_conc":256,"lower_bound":false,"samples":61157},{"bound_ms":100,"rps":{"value":10191,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2455,"first_disqualified_conc":1024,"lower_bound":false,"samples":61157},{"bound_ms":null,"rps":{"value":10191,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2455,"first_disqualified_conc":null,"lower_bound":false,"samples":61157}],"sweep":[{"conc":1,"ok":20765,"rps":3460,"p99_us":323,"fail":0},{"conc":1,"ok":20754,"rps":3458,"p99_us":322,"fail":0},{"conc":1,"ok":20802,"rps":3466,"p99_us":323,"fail":0},{"conc":2,"ok":34514,"rps":5751,"p99_us":432,"fail":0},{"conc":2,"ok":34641,"rps":5772,"p99_us":428,"fail":0},{"conc":2,"ok":34629,"rps":5770,"p99_us":429,"fail":0},{"conc":4,"ok":48312,"rps":8050,"p99_us":704,"fail":0},{"conc":4,"ok":48256,"rps":8041,"p99_us":701,"fail":0},{"conc":4,"ok":48309,"rps":8050,"p99_us":705,"fail":0},{"conc":8,"ok":61102,"rps":10182,"p99_us":1215,"fail":0},{"conc":8,"ok":61016,"rps":10168,"p99_us":1221,"fail":0},{"conc":8,"ok":61058,"rps":10175,"p99_us":1219,"fail":0},{"conc":16,"ok":61090,"rps":10180,"p99_us":2472,"fail":0},{"conc":16,"ok":61095,"rps":10181,"p99_us":2458,"fail":0},{"conc":16,"ok":61157,"rps":10191,"p99_us":2455,"fail":0},{"conc":32,"ok":60885,"rps":10146,"p99_us":5127,"fail":0},{"conc":32,"ok":60805,"rps":10133,"p99_us":5115,"fail":0},{"conc":32,"ok":60955,"rps":10158,"p99_us":5102,"fail":0},{"conc":64,"ok":60273,"rps":10044,"p99_us":10992,"fail":0},{"conc":64,"ok":60174,"rps":10027,"p99_us":11088,"fail":0},{"conc":64,"ok":60296,"rps":10047,"p99_us":10944,"fail":0},{"conc":128,"ok":60189,"rps":10030,"p99_us":24576,"fail":0},{"conc":128,"ok":60235,"rps":10037,"p99_us":24952,"fail":0},{"conc":128,"ok":60192,"rps":10030,"p99_us":24885,"fail":0},{"conc":256,"ok":59301,"rps":9882,"p99_us":56288,"fail":0},{"conc":256,"ok":59509,"rps":9916,"p99_us":54741,"fail":0},{"conc":256,"ok":59587,"rps":9930,"p99_us":58149,"fail":0},{"conc":512,"ok":58428,"rps":9737,"p99_us":93541,"fail":0},{"conc":512,"ok":58396,"rps":9732,"p99_us":93907,"fail":0},{"conc":512,"ok":58495,"rps":9747,"p99_us":94976,"fail":0},{"conc":1024,"ok":57512,"rps":9585,"p99_us":139994,"fail":0},{"conc":1024,"ok":57645,"rps":9605,"p99_us":130917,"fail":0},{"conc":1024,"ok":57654,"rps":9607,"p99_us":131393,"fail":0},{"conc":2048,"ok":56617,"rps":9435,"p99_us":382168,"fail":0},{"conc":2048,"ok":57035,"rps":9504,"p99_us":334615,"fail":0},{"conc":2048,"ok":56807,"rps":9466,"p99_us":311028,"fail":0},{"conc":4096,"ok":55823,"rps":9302,"p99_us":652752,"fail":0},{"conc":4096,"ok":56155,"rps":9357,"p99_us":925543,"fail":0},{"conc":4096,"ok":56330,"rps":9387,"p99_us":387617,"fail":0},{"conc":8192,"ok":54699,"rps":9115,"p99_us":987954,"fail":0},{"conc":8192,"ok":55440,"rps":9238,"p99_us":2116028,"fail":0},{"conc":8192,"ok":55818,"rps":9301,"p99_us":1881474,"fail":0},{"conc":16384,"ok":53796,"rps":8964,"p99_us":2909263,"fail":0},{"conc":16384,"ok":53283,"rps":8878,"p99_us":2077023,"fail":0},{"conc":16384,"ok":53565,"rps":8926,"p99_us":2103073,"fail":0},{"conc":32768,"ok":18770,"rps":3125,"p99_us":1238031,"fail":0},{"conc":32768,"ok":30518,"rps":5085,"p99_us":5767154,"fail":0},{"conc":32768,"ok":9074,"rps":1507,"p99_us":988646,"fail":0}],"rungs":[{"conc":1,"rps":{"value":3460,"certified":true,"suppressed":false},"p99_us":323,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":5770,"certified":true,"suppressed":false},"p99_us":429,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":8050,"certified":true,"suppressed":false},"p99_us":704,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":10175,"certified":true,"suppressed":false},"p99_us":1219,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":10181,"certified":true,"suppressed":false},"p99_us":2458,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":10146,"certified":true,"suppressed":false},"p99_us":5115,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":10044,"certified":true,"suppressed":false},"p99_us":10992,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":10030,"certified":true,"suppressed":false},"p99_us":24885,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":9916,"certified":true,"suppressed":false},"p99_us":56288,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":9737,"certified":true,"suppressed":false},"p99_us":93907,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":9605,"certified":true,"suppressed":false},"p99_us":131393,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":9466,"certified":true,"suppressed":false},"p99_us":334615,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":9357,"certified":true,"suppressed":false},"p99_us":652752,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":9238,"certified":true,"suppressed":false},"p99_us":1881474,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":8926,"certified":true,"suppressed":false},"p99_us":2103073,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":3125,"certified":true,"suppressed":false},"p99_us":1238031,"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 20661 successful gateway round trip(s) and 267569 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":20661,"direct_c1_samples":267569},"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":12.7421875,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":19.47265625,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":19.28515625,"certified":true,"suppressed":false},"peak_rss_mib":{"value":19.5,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":19.5,"certified":true,"suppressed":false},"plateaued":true,"time_to_plateau_s":{"value":66,"certified":true,"suppressed":false},"growth_rate_mib_per_min":{"value":0.028056430633918075,"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,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":13.33203125},{"t_s":2,"rss_mib":18.11328125},{"t_s":2,"rss_mib":18.11328125},{"t_s":5,"rss_mib":18.25390625},{"t_s":5,"rss_mib":18.25390625},{"t_s":8,"rss_mib":18.80859375},{"t_s":8,"rss_mib":18.80859375},{"t_s":10,"rss_mib":18.83203125},{"t_s":11,"rss_mib":18.83203125},{"t_s":14,"rss_mib":18.8828125},{"t_s":15,"rss_mib":18.8828125},{"t_s":16,"rss_mib":18.92578125},{"t_s":17,"rss_mib":18.92578125},{"t_s":20,"rss_mib":18.9921875},{"t_s":20,"rss_mib":18.9921875},{"t_s":24,"rss_mib":18.9921875},{"t_s":26,"rss_mib":18.9921875},{"t_s":29,"rss_mib":19},{"t_s":30,"rss_mib":19},{"t_s":33,"rss_mib":19},{"t_s":35,"rss_mib":19},{"t_s":39,"rss_mib":19},{"t_s":42,"rss_mib":19},{"t_s":43,"rss_mib":19.03125},{"t_s":45,"rss_mib":19.03125},{"t_s":46,"rss_mib":19.03515625},{"t_s":48,"rss_mib":19.03515625},{"t_s":50,"rss_mib":19.03515625},{"t_s":54,"rss_mib":19.03515625},{"t_s":57,"rss_mib":19.03515625},{"t_s":60,"rss_mib":19.03515625},{"t_s":63,"rss_mib":19.03515625},{"t_s":65,"rss_mib":19.0390625},{"t_s":66,"rss_mib":19.0390625},{"t_s":69,"rss_mib":19.0390625},{"t_s":71,"rss_mib":19.11328125},{"t_s":72,"rss_mib":19.11328125},{"t_s":75,"rss_mib":19.11328125},{"t_s":78,"rss_mib":19.11328125},{"t_s":78,"rss_mib":19.1171875},{"t_s":81,"rss_mib":19.1171875},{"t_s":84,"rss_mib":19.1171875},{"t_s":86,"rss_mib":19.125},{"t_s":87,"rss_mib":19.125},{"t_s":89,"rss_mib":19.12890625},{"t_s":90,"rss_mib":19.12890625},{"t_s":93,"rss_mib":19.12890625},{"t_s":96,"rss_mib":19.1328125},{"t_s":96,"rss_mib":19.21484375},{"t_s":99,"rss_mib":19.21484375},{"t_s":102,"rss_mib":19.21484375},{"t_s":105,"rss_mib":19.21484375},{"t_s":108,"rss_mib":19.21484375},{"t_s":111,"rss_mib":19.21484375},{"t_s":114,"rss_mib":19.21484375},{"t_s":117,"rss_mib":19.21484375},{"t_s":120,"rss_mib":19.21484375},{"t_s":123,"rss_mib":19.21484375},{"t_s":126,"rss_mib":19.21484375},{"t_s":129,"rss_mib":19.21484375},{"t_s":132,"rss_mib":19.21484375},{"t_s":135,"rss_mib":19.21484375},{"t_s":138,"rss_mib":19.21484375},{"t_s":141,"rss_mib":19.21484375},{"t_s":144,"rss_mib":19.22265625},{"t_s":144,"rss_mib":19.46484375},{"t_s":147,"rss_mib":19.46484375},{"t_s":150,"rss_mib":19.46484375},{"t_s":153,"rss_mib":19.46484375},{"t_s":156,"rss_mib":19.46484375},{"t_s":159,"rss_mib":19.46484375},{"t_s":162,"rss_mib":19.46484375},{"t_s":165,"rss_mib":19.46484375},{"t_s":168,"rss_mib":19.46484375},{"t_s":171,"rss_mib":19.46484375},{"t_s":173,"rss_mib":19.46875},{"t_s":174,"rss_mib":19.46875},{"t_s":177,"rss_mib":19.46875},{"t_s":180,"rss_mib":19.46875},{"t_s":180,"rss_mib":19.47265625},{"t_s":183,"rss_mib":19.47265625},{"t_s":186,"rss_mib":19.47265625},{"t_s":189,"rss_mib":19.47265625},{"t_s":192,"rss_mib":19.47265625},{"t_s":195,"rss_mib":19.47265625},{"t_s":198,"rss_mib":19.47265625},{"t_s":201,"rss_mib":19.47265625},{"t_s":204,"rss_mib":19.47265625},{"t_s":207,"rss_mib":19.47265625},{"t_s":210,"rss_mib":19.47265625},{"t_s":213,"rss_mib":19.47265625},{"t_s":216,"rss_mib":19.47265625},{"t_s":219,"rss_mib":19.47265625},{"t_s":222,"rss_mib":19.47265625},{"t_s":225,"rss_mib":19.47265625},{"t_s":228,"rss_mib":19.47265625},{"t_s":231,"rss_mib":19.47265625},{"t_s":234,"rss_mib":19.47265625},{"t_s":237,"rss_mib":19.47265625},{"t_s":240,"rss_mib":19.47265625},{"t_s":243,"rss_mib":19.47265625},{"t_s":246,"rss_mib":19.47265625},{"t_s":249,"rss_mib":19.48046875},{"t_s":249,"rss_mib":19.48046875},{"t_s":252,"rss_mib":19.48046875},{"t_s":255,"rss_mib":19.48046875},{"t_s":258,"rss_mib":19.48046875},{"t_s":261,"rss_mib":19.48046875},{"t_s":264,"rss_mib":19.48046875},{"t_s":267,"rss_mib":19.48046875},{"t_s":270,"rss_mib":19.48046875},{"t_s":273,"rss_mib":19.48046875},{"t_s":276,"rss_mib":19.48046875},{"t_s":279,"rss_mib":19.48046875},{"t_s":282,"rss_mib":19.48046875},{"t_s":285,"rss_mib":19.5},{"t_s":285,"rss_mib":19.5},{"t_s":288,"rss_mib":19.5},{"t_s":291,"rss_mib":19.5},{"t_s":294,"rss_mib":19.5},{"t_s":297,"rss_mib":19.5},{"t_s":300,"rss_mib":19.5},{"t_s":303,"rss_mib":19.5},{"t_s":306,"rss_mib":19.5},{"t_s":309,"rss_mib":19.5},{"t_s":312,"rss_mib":19.5},{"t_s":315,"rss_mib":19.5},{"t_s":318,"rss_mib":19.5},{"t_s":321,"rss_mib":19.5},{"t_s":324,"rss_mib":19.5},{"t_s":327,"rss_mib":19.5},{"t_s":329,"rss_mib":19.28515625},{"t_s":330,"rss_mib":19.28515625},{"t_s":333,"rss_mib":19.28515625},{"t_s":336,"rss_mib":19.28515625},{"t_s":339,"rss_mib":19.28515625},{"t_s":342,"rss_mib":19.28515625},{"t_s":345,"rss_mib":19.28515625},{"t_s":348,"rss_mib":19.28515625},{"t_s":351,"rss_mib":19.28515625},{"t_s":354,"rss_mib":19.28515625},{"t_s":357,"rss_mib":19.28515625},{"t_s":360,"rss_mib":19.28515625}],"idle_rss_series":[{"t_s":0,"rss_mib":12.7421875},{"t_s":0,"rss_mib":12.7421875},{"t_s":0,"rss_mib":12.7421875},{"t_s":1,"rss_mib":12.7421875},{"t_s":1,"rss_mib":12.7421875},{"t_s":2,"rss_mib":12.7421875},{"t_s":2,"rss_mib":12.7421875},{"t_s":3,"rss_mib":12.7421875},{"t_s":3,"rss_mib":12.7421875},{"t_s":4,"rss_mib":12.7421875},{"t_s":5,"rss_mib":12.7421875},{"t_s":5,"rss_mib":12.7421875},{"t_s":5,"rss_mib":12.7421875},{"t_s":6,"rss_mib":12.7421875},{"t_s":6,"rss_mib":12.7421875},{"t_s":7,"rss_mib":12.7421875},{"t_s":7,"rss_mib":12.7421875},{"t_s":8,"rss_mib":12.7421875},{"t_s":8,"rss_mib":12.7421875},{"t_s":9,"rss_mib":12.7421875},{"t_s":10,"rss_mib":12.7421875},{"t_s":10,"rss_mib":12.7421875},{"t_s":10,"rss_mib":12.7421875},{"t_s":11,"rss_mib":12.7421875},{"t_s":11,"rss_mib":12.7421875},{"t_s":12,"rss_mib":12.7421875},{"t_s":12,"rss_mib":12.7421875},{"t_s":13,"rss_mib":12.7421875},{"t_s":13,"rss_mib":12.7421875},{"t_s":14,"rss_mib":12.7421875},{"t_s":15,"rss_mib":12.7421875},{"t_s":15,"rss_mib":12.7421875},{"t_s":15,"rss_mib":12.7421875},{"t_s":16,"rss_mib":12.7421875},{"t_s":16,"rss_mib":12.7421875},{"t_s":17,"rss_mib":12.7421875},{"t_s":17,"rss_mib":12.7421875},{"t_s":18,"rss_mib":12.7421875},{"t_s":18,"rss_mib":12.7421875},{"t_s":19,"rss_mib":12.7421875},{"t_s":20,"rss_mib":12.7421875},{"t_s":20,"rss_mib":12.7421875},{"t_s":20,"rss_mib":12.7421875},{"t_s":21,"rss_mib":12.7421875},{"t_s":21,"rss_mib":12.7421875},{"t_s":22,"rss_mib":12.7421875},{"t_s":22,"rss_mib":12.7421875},{"t_s":23,"rss_mib":12.7421875},{"t_s":23,"rss_mib":12.7421875},{"t_s":24,"rss_mib":12.7421875},{"t_s":25,"rss_mib":12.7421875},{"t_s":25,"rss_mib":12.7421875},{"t_s":25,"rss_mib":12.7421875},{"t_s":26,"rss_mib":12.7421875},{"t_s":26,"rss_mib":12.7421875},{"t_s":27,"rss_mib":12.7421875},{"t_s":27,"rss_mib":12.7421875},{"t_s":28,"rss_mib":12.7421875},{"t_s":28,"rss_mib":12.7421875},{"t_s":29,"rss_mib":12.7421875},{"t_s":30,"rss_mib":12.7421875},{"t_s":30,"rss_mib":12.7421875},{"t_s":30,"rss_mib":12.7421875},{"t_s":31,"rss_mib":12.7421875},{"t_s":31,"rss_mib":12.7421875},{"t_s":32,"rss_mib":12.7421875},{"t_s":32,"rss_mib":12.7421875},{"t_s":33,"rss_mib":12.7421875},{"t_s":33,"rss_mib":12.7421875},{"t_s":34,"rss_mib":12.7421875},{"t_s":35,"rss_mib":12.7421875},{"t_s":35,"rss_mib":12.7421875},{"t_s":35,"rss_mib":12.7421875},{"t_s":36,"rss_mib":12.7421875},{"t_s":36,"rss_mib":12.7421875},{"t_s":37,"rss_mib":12.7421875},{"t_s":37,"rss_mib":12.7421875},{"t_s":38,"rss_mib":12.7421875},{"t_s":38,"rss_mib":12.7421875},{"t_s":39,"rss_mib":12.7421875},{"t_s":40,"rss_mib":12.7421875},{"t_s":40,"rss_mib":12.7421875},{"t_s":40,"rss_mib":12.7421875},{"t_s":41,"rss_mib":12.7421875},{"t_s":41,"rss_mib":12.7421875},{"t_s":42,"rss_mib":12.7421875},{"t_s":42,"rss_mib":12.7421875},{"t_s":43,"rss_mib":12.7421875},{"t_s":43,"rss_mib":12.7421875},{"t_s":44,"rss_mib":12.7421875},{"t_s":45,"rss_mib":12.7421875},{"t_s":45,"rss_mib":12.7421875},{"t_s":45,"rss_mib":12.7421875},{"t_s":46,"rss_mib":12.7421875},{"t_s":46,"rss_mib":12.7421875},{"t_s":47,"rss_mib":12.7421875},{"t_s":47,"rss_mib":12.7421875},{"t_s":48,"rss_mib":12.7421875},{"t_s":48,"rss_mib":12.7421875},{"t_s":49,"rss_mib":12.7421875},{"t_s":50,"rss_mib":12.7421875},{"t_s":50,"rss_mib":12.7421875},{"t_s":50,"rss_mib":12.7421875},{"t_s":51,"rss_mib":12.7421875},{"t_s":51,"rss_mib":12.7421875},{"t_s":52,"rss_mib":12.7421875},{"t_s":52,"rss_mib":12.7421875},{"t_s":53,"rss_mib":12.7421875},{"t_s":53,"rss_mib":12.7421875},{"t_s":54,"rss_mib":12.7421875},{"t_s":55,"rss_mib":12.7421875},{"t_s":55,"rss_mib":12.7421875},{"t_s":55,"rss_mib":12.7421875},{"t_s":56,"rss_mib":12.7421875},{"t_s":56,"rss_mib":12.7421875},{"t_s":57,"rss_mib":12.7421875},{"t_s":57,"rss_mib":12.7421875},{"t_s":58,"rss_mib":12.7421875},{"t_s":58,"rss_mib":12.7421875},{"t_s":59,"rss_mib":12.7421875},{"t_s":59,"rss_mib":12.7421875}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.008994143,"cost":6.026715669,"memory":421.883146339,"streaming":0.000001839,"streams_sustained":0.000001702,"throughput":606.763066411},"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"}}},"openai":{"served":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/chat/completions","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/messages","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":{"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":253,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":283,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":311,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":277.0705901801951,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":3609.1885441527443,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.6965431268306815,"certified":true,"suppressed":false},"cost_window_ok":{"value":60490,"certified":true,"suppressed":false},"cost_window_rps":{"value":10080,"certified":true,"suppressed":false},"cost_threads":{"value":5,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":8572,"certified":true,"suppressed":false},"concurrency":4,"p99_us":594,"first_disqualified_conc":8,"lower_bound":false,"samples":51441},{"bound_ms":5,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":32,"lower_bound":false,"samples":60500},{"bound_ms":10,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":64,"lower_bound":false,"samples":60500},{"bound_ms":50,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":256,"lower_bound":false,"samples":60500},{"bound_ms":100,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":1024,"lower_bound":false,"samples":60500},{"bound_ms":null,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":null,"lower_bound":false,"samples":60500}],"sweep":[{"conc":1,"ok":21640,"rps":3606,"p99_us":311,"fail":0},{"conc":1,"ok":21650,"rps":3607,"p99_us":311,"fail":0},{"conc":1,"ok":21569,"rps":3594,"p99_us":312,"fail":0},{"conc":2,"ok":31855,"rps":5308,"p99_us":418,"fail":0},{"conc":2,"ok":31701,"rps":5282,"p99_us":419,"fail":0},{"conc":2,"ok":31644,"rps":5273,"p99_us":420,"fail":0},{"conc":4,"ok":51006,"rps":8499,"p99_us":595,"fail":0},{"conc":4,"ok":51441,"rps":8572,"p99_us":594,"fail":0},{"conc":4,"ok":51316,"rps":8551,"p99_us":593,"fail":0},{"conc":8,"ok":60286,"rps":10046,"p99_us":1203,"fail":0},{"conc":8,"ok":60318,"rps":10051,"p99_us":1209,"fail":0},{"conc":8,"ok":60327,"rps":10053,"p99_us":1209,"fail":0},{"conc":16,"ok":60324,"rps":10052,"p99_us":2493,"fail":0},{"conc":16,"ok":60500,"rps":10082,"p99_us":2482,"fail":0},{"conc":16,"ok":60325,"rps":10053,"p99_us":2496,"fail":0},{"conc":32,"ok":59850,"rps":9974,"p99_us":5218,"fail":0},{"conc":32,"ok":59868,"rps":9977,"p99_us":5187,"fail":0},{"conc":32,"ok":59897,"rps":9982,"p99_us":5199,"fail":0},{"conc":64,"ok":59227,"rps":9870,"p99_us":11146,"fail":0},{"conc":64,"ok":59387,"rps":9897,"p99_us":11181,"fail":0},{"conc":64,"ok":59404,"rps":9900,"p99_us":11163,"fail":0},{"conc":128,"ok":59602,"rps":9932,"p99_us":25187,"fail":0},{"conc":128,"ok":59509,"rps":9916,"p99_us":25168,"fail":0},{"conc":128,"ok":59427,"rps":9904,"p99_us":24757,"fail":0},{"conc":256,"ok":59085,"rps":9846,"p99_us":55035,"fail":0},{"conc":256,"ok":59251,"rps":9874,"p99_us":57249,"fail":0},{"conc":256,"ok":59368,"rps":9893,"p99_us":57642,"fail":0},{"conc":512,"ok":58011,"rps":9668,"p99_us":94197,"fail":0},{"conc":512,"ok":58049,"rps":9674,"p99_us":93858,"fail":0},{"conc":512,"ok":58273,"rps":9711,"p99_us":94318,"fail":0},{"conc":1024,"ok":56680,"rps":9445,"p99_us":138751,"fail":0},{"conc":1024,"ok":57008,"rps":9499,"p99_us":137676,"fail":0},{"conc":1024,"ok":56798,"rps":9464,"p99_us":132484,"fail":0},{"conc":2048,"ok":55560,"rps":9258,"p99_us":328520,"fail":0},{"conc":2048,"ok":55986,"rps":9329,"p99_us":367636,"fail":0},{"conc":2048,"ok":56137,"rps":9354,"p99_us":340707,"fail":0},{"conc":4096,"ok":55199,"rps":9199,"p99_us":854765,"fail":0},{"conc":4096,"ok":56235,"rps":9372,"p99_us":465530,"fail":0},{"conc":4096,"ok":54901,"rps":9148,"p99_us":1368193,"fail":0},{"conc":8192,"ok":54255,"rps":9041,"p99_us":2169098,"fail":0},{"conc":8192,"ok":53775,"rps":8961,"p99_us":2316246,"fail":0},{"conc":8192,"ok":54267,"rps":9043,"p99_us":1786177,"fail":0},{"conc":16384,"ok":52209,"rps":8700,"p99_us":2878906,"fail":0},{"conc":16384,"ok":52949,"rps":8824,"p99_us":2916402,"fail":0},{"conc":16384,"ok":52168,"rps":8693,"p99_us":2634896,"fail":0},{"conc":32768,"ok":16730,"rps":2780,"p99_us":1164793,"fail":0},{"conc":32768,"ok":32916,"rps":5485,"p99_us":5439856,"fail":0},{"conc":32768,"ok":32440,"rps":5405,"p99_us":3847695,"fail":0}],"rungs":[{"conc":1,"rps":{"value":3606,"certified":true,"suppressed":false},"p99_us":311,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":5282,"certified":true,"suppressed":false},"p99_us":419,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":8551,"certified":true,"suppressed":false},"p99_us":594,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":10051,"certified":true,"suppressed":false},"p99_us":1209,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":10053,"certified":true,"suppressed":false},"p99_us":2493,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":9977,"certified":true,"suppressed":false},"p99_us":5199,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":9897,"certified":true,"suppressed":false},"p99_us":11163,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":9916,"certified":true,"suppressed":false},"p99_us":25168,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":9874,"certified":true,"suppressed":false},"p99_us":57249,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":9674,"certified":true,"suppressed":false},"p99_us":94197,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":9464,"certified":true,"suppressed":false},"p99_us":137676,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":9329,"certified":true,"suppressed":false},"p99_us":340707,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":9199,"certified":true,"suppressed":false},"p99_us":854765,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":9041,"certified":true,"suppressed":false},"p99_us":2169098,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":8700,"certified":true,"suppressed":false},"p99_us":2878906,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":5405,"certified":true,"suppressed":false},"p99_us":3847695,"windows":3,"clean_windows":3}],"egress_reverified":true,"c1_note":"the c=1 percentiles are taken over 21651 successful gateway round trip(s) and 262942 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":21651,"direct_c1_samples":262942},"stream":{"stream_served":"not_measured","stream_error":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","reason":"not_measured","stream_c1_note":null,"added_ttft_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"added_ttft_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"added_gap_p50_us":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"added_gap_p99_us":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"streams_sustained_fps":{"value":0,"certified":true,"suppressed":false,"note":"measured_failure"},"streams_sustained":{"value":0,"certified":true,"suppressed":false,"note":"measured_failure"}},"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":12.7578125,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":19.046875,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":18.984375,"certified":true,"suppressed":false},"peak_rss_mib":{"value":19.66015625,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":19.33203125,"certified":true,"suppressed":false},"plateaued":false,"time_to_plateau_s":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"memory was still oscillating with no net trend - it returns to where it was, which is not a leak when the 300s load window closed, so it has no time-to-plateau; its rate is published beside this"},"growth_rate_mib_per_min":{"value":0.3275657622729096,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":0,"certified":true,"suppressed":false},"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,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":12.90234375},{"t_s":2,"rss_mib":18.00390625},{"t_s":2,"rss_mib":18.00390625},{"t_s":5,"rss_mib":18.2890625},{"t_s":5,"rss_mib":18.2890625},{"t_s":8,"rss_mib":18.50390625},{"t_s":8,"rss_mib":18.50390625},{"t_s":9,"rss_mib":18.51171875},{"t_s":11,"rss_mib":18.51171875},{"t_s":13,"rss_mib":18.53125},{"t_s":15,"rss_mib":18.55078125},{"t_s":16,"rss_mib":18.63671875},{"t_s":17,"rss_mib":18.63671875},{"t_s":19,"rss_mib":18.66015625},{"t_s":20,"rss_mib":18.66015625},{"t_s":24,"rss_mib":18.66015625},{"t_s":26,"rss_mib":19.515625},{"t_s":26,"rss_mib":19.515625},{"t_s":29,"rss_mib":19.5546875},{"t_s":30,"rss_mib":19.5546875},{"t_s":30,"rss_mib":19.125},{"t_s":32,"rss_mib":19.13671875},{"t_s":35,"rss_mib":19.15234375},{"t_s":35,"rss_mib":19.15234375},{"t_s":36,"rss_mib":18.765625},{"t_s":39,"rss_mib":18.765625},{"t_s":41,"rss_mib":18.765625},{"t_s":45,"rss_mib":18.765625},{"t_s":48,"rss_mib":18.765625},{"t_s":50,"rss_mib":18.765625},{"t_s":54,"rss_mib":18.765625},{"t_s":55,"rss_mib":18.86328125},{"t_s":57,"rss_mib":18.86328125},{"t_s":60,"rss_mib":18.86328125},{"t_s":63,"rss_mib":18.86328125},{"t_s":65,"rss_mib":18.86328125},{"t_s":66,"rss_mib":18.8671875},{"t_s":69,"rss_mib":18.8671875},{"t_s":72,"rss_mib":18.8671875},{"t_s":72,"rss_mib":19.16015625},{"t_s":75,"rss_mib":19.16015625},{"t_s":78,"rss_mib":19.16015625},{"t_s":78,"rss_mib":19.16796875},{"t_s":81,"rss_mib":19.16796875},{"t_s":84,"rss_mib":19.16796875},{"t_s":84,"rss_mib":18.69921875},{"t_s":87,"rss_mib":18.83203125},{"t_s":90,"rss_mib":18.83203125},{"t_s":90,"rss_mib":18.8359375},{"t_s":93,"rss_mib":18.8359375},{"t_s":96,"rss_mib":18.8359375},{"t_s":99,"rss_mib":18.8359375},{"t_s":102,"rss_mib":18.8359375},{"t_s":104,"rss_mib":18.83984375},{"t_s":105,"rss_mib":18.83984375},{"t_s":108,"rss_mib":18.83984375},{"t_s":108,"rss_mib":19.03125},{"t_s":111,"rss_mib":19.03125},{"t_s":111,"rss_mib":19.0390625},{"t_s":114,"rss_mib":19.0390625},{"t_s":114,"rss_mib":19.47265625},{"t_s":117,"rss_mib":19.47265625},{"t_s":120,"rss_mib":19.47265625},{"t_s":120,"rss_mib":19.16796875},{"t_s":123,"rss_mib":19.16796875},{"t_s":124,"rss_mib":18.88671875},{"t_s":126,"rss_mib":18.88671875},{"t_s":126,"rss_mib":18.6640625},{"t_s":129,"rss_mib":18.6875},{"t_s":131,"rss_mib":18.73828125},{"t_s":132,"rss_mib":18.73828125},{"t_s":135,"rss_mib":18.73828125},{"t_s":138,"rss_mib":18.73828125},{"t_s":139,"rss_mib":19.15234375},{"t_s":141,"rss_mib":19.15234375},{"t_s":144,"rss_mib":19.15234375},{"t_s":144,"rss_mib":19.50390625},{"t_s":147,"rss_mib":19.50390625},{"t_s":150,"rss_mib":19.50390625},{"t_s":150,"rss_mib":19.546875},{"t_s":153,"rss_mib":19.546875},{"t_s":156,"rss_mib":19.546875},{"t_s":159,"rss_mib":19.546875},{"t_s":160,"rss_mib":19.26171875},{"t_s":162,"rss_mib":19.26171875},{"t_s":162,"rss_mib":18.67578125},{"t_s":165,"rss_mib":18.6796875},{"t_s":168,"rss_mib":18.6796875},{"t_s":170,"rss_mib":18.75},{"t_s":171,"rss_mib":18.75},{"t_s":174,"rss_mib":18.75},{"t_s":174,"rss_mib":18.9296875},{"t_s":177,"rss_mib":18.9296875},{"t_s":180,"rss_mib":18.9296875},{"t_s":183,"rss_mib":18.9296875},{"t_s":186,"rss_mib":18.9296875},{"t_s":186,"rss_mib":19.66015625},{"t_s":189,"rss_mib":19.66015625},{"t_s":192,"rss_mib":19.66015625},{"t_s":192,"rss_mib":19.4609375},{"t_s":195,"rss_mib":19.46875},{"t_s":198,"rss_mib":19.46875},{"t_s":201,"rss_mib":19.46875},{"t_s":201,"rss_mib":19.47265625},{"t_s":204,"rss_mib":19.47265625},{"t_s":204,"rss_mib":19.4765625},{"t_s":207,"rss_mib":19.4765625},{"t_s":209,"rss_mib":19.48046875},{"t_s":210,"rss_mib":19.48046875},{"t_s":213,"rss_mib":19.48046875},{"t_s":214,"rss_mib":18.87109375},{"t_s":216,"rss_mib":18.87109375},{"t_s":216,"rss_mib":18.8984375},{"t_s":219,"rss_mib":18.8984375},{"t_s":220,"rss_mib":18.97265625},{"t_s":223,"rss_mib":19.046875},{"t_s":223,"rss_mib":18.8203125},{"t_s":225,"rss_mib":18.82421875},{"t_s":225,"rss_mib":18.828125},{"t_s":228,"rss_mib":18.828125},{"t_s":229,"rss_mib":18.8359375},{"t_s":231,"rss_mib":18.8359375},{"t_s":233,"rss_mib":18.85546875},{"t_s":234,"rss_mib":18.85546875},{"t_s":237,"rss_mib":18.85546875},{"t_s":239,"rss_mib":18.86328125},{"t_s":240,"rss_mib":18.86328125},{"t_s":242,"rss_mib":18.90234375},{"t_s":243,"rss_mib":18.90234375},{"t_s":246,"rss_mib":18.90234375},{"t_s":249,"rss_mib":18.90234375},{"t_s":252,"rss_mib":18.90234375},{"t_s":255,"rss_mib":18.90234375},{"t_s":255,"rss_mib":18.9140625},{"t_s":258,"rss_mib":18.9140625},{"t_s":258,"rss_mib":19.5},{"t_s":261,"rss_mib":19.5},{"t_s":264,"rss_mib":19.5},{"t_s":264,"rss_mib":19.51171875},{"t_s":267,"rss_mib":19.51171875},{"t_s":270,"rss_mib":19.51171875},{"t_s":270,"rss_mib":19.2265625},{"t_s":273,"rss_mib":19.2265625},{"t_s":276,"rss_mib":19.2265625},{"t_s":278,"rss_mib":19.546875},{"t_s":279,"rss_mib":19.546875},{"t_s":282,"rss_mib":19.546875},{"t_s":282,"rss_mib":19.328125},{"t_s":285,"rss_mib":19.328125},{"t_s":288,"rss_mib":19.328125},{"t_s":291,"rss_mib":19.328125},{"t_s":293,"rss_mib":18.9453125},{"t_s":294,"rss_mib":18.9453125},{"t_s":295,"rss_mib":18.95703125},{"t_s":297,"rss_mib":18.95703125},{"t_s":297,"rss_mib":18.984375},{"t_s":300,"rss_mib":18.984375},{"t_s":303,"rss_mib":18.984375},{"t_s":306,"rss_mib":18.984375},{"t_s":309,"rss_mib":18.984375},{"t_s":312,"rss_mib":18.984375},{"t_s":315,"rss_mib":18.984375},{"t_s":318,"rss_mib":18.984375},{"t_s":321,"rss_mib":18.984375},{"t_s":324,"rss_mib":18.984375},{"t_s":327,"rss_mib":18.984375},{"t_s":330,"rss_mib":18.984375},{"t_s":333,"rss_mib":18.984375},{"t_s":336,"rss_mib":18.984375},{"t_s":339,"rss_mib":18.984375},{"t_s":342,"rss_mib":18.984375},{"t_s":345,"rss_mib":18.984375},{"t_s":348,"rss_mib":18.984375},{"t_s":351,"rss_mib":18.984375},{"t_s":354,"rss_mib":18.984375},{"t_s":357,"rss_mib":18.984375},{"t_s":360,"rss_mib":18.984375}],"idle_rss_series":[{"t_s":0,"rss_mib":12.7578125},{"t_s":0,"rss_mib":12.7578125},{"t_s":0,"rss_mib":12.7578125},{"t_s":1,"rss_mib":12.7578125},{"t_s":1,"rss_mib":12.7578125},{"t_s":2,"rss_mib":12.7578125},{"t_s":2,"rss_mib":12.7578125},{"t_s":3,"rss_mib":12.7578125},{"t_s":3,"rss_mib":12.7578125},{"t_s":4,"rss_mib":12.7578125},{"t_s":5,"rss_mib":12.7578125},{"t_s":5,"rss_mib":12.7578125},{"t_s":5,"rss_mib":12.7578125},{"t_s":6,"rss_mib":12.7578125},{"t_s":6,"rss_mib":12.7578125},{"t_s":7,"rss_mib":12.7578125},{"t_s":7,"rss_mib":12.7578125},{"t_s":8,"rss_mib":12.7578125},{"t_s":8,"rss_mib":12.7578125},{"t_s":9,"rss_mib":12.7578125},{"t_s":10,"rss_mib":12.7578125},{"t_s":10,"rss_mib":12.7578125},{"t_s":10,"rss_mib":12.7578125},{"t_s":11,"rss_mib":12.7578125},{"t_s":11,"rss_mib":12.7578125},{"t_s":12,"rss_mib":12.7578125},{"t_s":12,"rss_mib":12.7578125},{"t_s":13,"rss_mib":12.7578125},{"t_s":13,"rss_mib":12.7578125},{"t_s":14,"rss_mib":12.7578125},{"t_s":15,"rss_mib":12.7578125},{"t_s":15,"rss_mib":12.7578125},{"t_s":15,"rss_mib":12.7578125},{"t_s":16,"rss_mib":12.7578125},{"t_s":16,"rss_mib":12.7578125},{"t_s":17,"rss_mib":12.7578125},{"t_s":17,"rss_mib":12.7578125},{"t_s":18,"rss_mib":12.7578125},{"t_s":18,"rss_mib":12.7578125},{"t_s":19,"rss_mib":12.7578125},{"t_s":20,"rss_mib":12.7578125},{"t_s":20,"rss_mib":12.7578125},{"t_s":20,"rss_mib":12.7578125},{"t_s":21,"rss_mib":12.7578125},{"t_s":21,"rss_mib":12.7578125},{"t_s":22,"rss_mib":12.7578125},{"t_s":22,"rss_mib":12.7578125},{"t_s":23,"rss_mib":12.7578125},{"t_s":23,"rss_mib":12.7578125},{"t_s":24,"rss_mib":12.7578125},{"t_s":25,"rss_mib":12.7578125},{"t_s":25,"rss_mib":12.7578125},{"t_s":25,"rss_mib":12.7578125},{"t_s":26,"rss_mib":12.7578125},{"t_s":26,"rss_mib":12.7578125},{"t_s":27,"rss_mib":12.7578125},{"t_s":27,"rss_mib":12.7578125},{"t_s":28,"rss_mib":12.7578125},{"t_s":28,"rss_mib":12.7578125},{"t_s":29,"rss_mib":12.7578125},{"t_s":30,"rss_mib":12.7578125},{"t_s":30,"rss_mib":12.7578125},{"t_s":30,"rss_mib":12.7578125},{"t_s":31,"rss_mib":12.7578125},{"t_s":31,"rss_mib":12.7578125},{"t_s":32,"rss_mib":12.7578125},{"t_s":32,"rss_mib":12.7578125},{"t_s":33,"rss_mib":12.7578125},{"t_s":33,"rss_mib":12.7578125},{"t_s":34,"rss_mib":12.7578125},{"t_s":35,"rss_mib":12.7578125},{"t_s":35,"rss_mib":12.7578125},{"t_s":35,"rss_mib":12.7578125},{"t_s":36,"rss_mib":12.7578125},{"t_s":36,"rss_mib":12.7578125},{"t_s":37,"rss_mib":12.7578125},{"t_s":37,"rss_mib":12.7578125},{"t_s":38,"rss_mib":12.7578125},{"t_s":38,"rss_mib":12.7578125},{"t_s":39,"rss_mib":12.7578125},{"t_s":40,"rss_mib":12.7578125},{"t_s":40,"rss_mib":12.7578125},{"t_s":40,"rss_mib":12.7578125},{"t_s":41,"rss_mib":12.7578125},{"t_s":41,"rss_mib":12.7578125},{"t_s":42,"rss_mib":12.7578125},{"t_s":42,"rss_mib":12.7578125},{"t_s":43,"rss_mib":12.7578125},{"t_s":43,"rss_mib":12.7578125},{"t_s":44,"rss_mib":12.7578125},{"t_s":45,"rss_mib":12.7578125},{"t_s":45,"rss_mib":12.7578125},{"t_s":45,"rss_mib":12.7578125},{"t_s":46,"rss_mib":12.7578125},{"t_s":46,"rss_mib":12.7578125},{"t_s":47,"rss_mib":12.7578125},{"t_s":47,"rss_mib":12.7578125},{"t_s":48,"rss_mib":12.7578125},{"t_s":48,"rss_mib":12.7578125},{"t_s":49,"rss_mib":12.7578125},{"t_s":50,"rss_mib":12.7578125},{"t_s":50,"rss_mib":12.7578125},{"t_s":50,"rss_mib":12.7578125},{"t_s":51,"rss_mib":12.7578125},{"t_s":51,"rss_mib":12.7578125},{"t_s":52,"rss_mib":12.7578125},{"t_s":52,"rss_mib":12.7578125},{"t_s":53,"rss_mib":12.7578125},{"t_s":53,"rss_mib":12.7578125},{"t_s":54,"rss_mib":12.7578125},{"t_s":55,"rss_mib":12.7578125},{"t_s":55,"rss_mib":12.7578125},{"t_s":55,"rss_mib":12.7578125},{"t_s":56,"rss_mib":12.7578125},{"t_s":56,"rss_mib":12.7578125},{"t_s":57,"rss_mib":12.7578125},{"t_s":57,"rss_mib":12.7578125},{"t_s":58,"rss_mib":12.7578125},{"t_s":59,"rss_mib":12.7578125},{"t_s":59,"rss_mib":12.7578125},{"t_s":59,"rss_mib":12.7578125}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.009170627,"cost":6.026211914,"memory":422.651899977,"streaming":1.304061042,"streams_sustained":0.001995084,"throughput":588.292474995},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.time_to_plateau_s":{"reason":"not_measured","detail":"memory was still oscillating with no net trend - it returns to where it was, which is not a leak when the 300s load window closed, so it has no time-to-plateau; its rate is published beside this"},"stream.added_gap_p50_us":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.added_gap_p99_us":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.added_ttft_p50_us":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.added_ttft_p99_us":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.gap_direct_samples":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.gap_gw_samples":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.streams_sustained_headroom":{"reason":"not_measured","detail":"the sustained figure was measured but the rig could not derive the mock's frame ceiling for this cell, so there is nothing to state this as a fraction of"},"stream.streams_sustained_mock_ceiling":{"reason":"not_measured","detail":"the sustained figure was measured but the rig could not derive the mock's frame ceiling for this cell, so there is nothing to state this as a fraction of"},"stream.ttft_direct_samples":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"},"stream.ttft_gw_samples":{"reason":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time"}}},"cohere":{"served":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v2/chat","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/model/bench-anthropic/converse","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1beta/models/bench-anthropic:generateContent","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":275,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":304,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":332,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":297.0902986640554,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":3365.97998822837,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.7061370243602647,"certified":true,"suppressed":false},"cost_window_ok":{"value":57188,"certified":true,"suppressed":false},"cost_window_rps":{"value":9530,"certified":true,"suppressed":false},"cost_threads":{"value":5,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":8878,"certified":true,"suppressed":false},"concurrency":4,"p99_us":606,"first_disqualified_conc":8,"lower_bound":false,"samples":53276},{"bound_ms":5,"rps":{"value":9865,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2555,"first_disqualified_conc":32,"lower_bound":false,"samples":59196},{"bound_ms":10,"rps":{"value":9865,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2555,"first_disqualified_conc":64,"lower_bound":false,"samples":59196},{"bound_ms":50,"rps":{"value":9865,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2555,"first_disqualified_conc":256,"lower_bound":false,"samples":59196},{"bound_ms":100,"rps":{"value":9865,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2555,"first_disqualified_conc":1024,"lower_bound":false,"samples":59196},{"bound_ms":null,"rps":{"value":9865,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2555,"first_disqualified_conc":null,"lower_bound":false,"samples":59196}],"sweep":[{"conc":1,"ok":20378,"rps":3395,"p99_us":327,"fail":0},{"conc":1,"ok":20374,"rps":3395,"p99_us":326,"fail":0},{"conc":1,"ok":20466,"rps":3410,"p99_us":326,"fail":0},{"conc":2,"ok":31894,"rps":5315,"p99_us":414,"fail":0},{"conc":2,"ok":31730,"rps":5287,"p99_us":416,"fail":0},{"conc":2,"ok":31737,"rps":5288,"p99_us":415,"fail":0},{"conc":4,"ok":53262,"rps":8875,"p99_us":603,"fail":0},{"conc":4,"ok":53015,"rps":8834,"p99_us":607,"fail":0},{"conc":4,"ok":53276,"rps":8878,"p99_us":606,"fail":0},{"conc":8,"ok":58900,"rps":9815,"p99_us":1222,"fail":0},{"conc":8,"ok":58954,"rps":9824,"p99_us":1214,"fail":0},{"conc":8,"ok":59063,"rps":9842,"p99_us":1213,"fail":0},{"conc":16,"ok":59196,"rps":9865,"p99_us":2555,"fail":0},{"conc":16,"ok":59029,"rps":9837,"p99_us":2563,"fail":0},{"conc":16,"ok":59035,"rps":9838,"p99_us":2557,"fail":0},{"conc":32,"ok":58732,"rps":9787,"p99_us":5296,"fail":0},{"conc":32,"ok":58529,"rps":9754,"p99_us":5318,"fail":0},{"conc":32,"ok":58703,"rps":9783,"p99_us":5302,"fail":0},{"conc":64,"ok":58169,"rps":9694,"p99_us":11281,"fail":0},{"conc":64,"ok":58407,"rps":9732,"p99_us":11261,"fail":0},{"conc":64,"ok":58259,"rps":9709,"p99_us":11300,"fail":0},{"conc":128,"ok":58984,"rps":9829,"p99_us":24743,"fail":0},{"conc":128,"ok":59123,"rps":9852,"p99_us":24919,"fail":0},{"conc":128,"ok":58932,"rps":9820,"p99_us":24831,"fail":0},{"conc":256,"ok":58930,"rps":9820,"p99_us":55818,"fail":0},{"conc":256,"ok":59145,"rps":9856,"p99_us":55464,"fail":0},{"conc":256,"ok":59123,"rps":9852,"p99_us":56044,"fail":0},{"conc":512,"ok":57694,"rps":9615,"p99_us":94769,"fail":0},{"conc":512,"ok":57778,"rps":9629,"p99_us":95586,"fail":0},{"conc":512,"ok":57805,"rps":9632,"p99_us":95520,"fail":0},{"conc":1024,"ok":55665,"rps":9277,"p99_us":143955,"fail":0},{"conc":1024,"ok":55610,"rps":9266,"p99_us":136025,"fail":0},{"conc":1024,"ok":55974,"rps":9327,"p99_us":144795,"fail":0},{"conc":2048,"ok":55431,"rps":9237,"p99_us":435792,"fail":0},{"conc":2048,"ok":54539,"rps":9088,"p99_us":325897,"fail":0},{"conc":2048,"ok":55147,"rps":9190,"p99_us":367277,"fail":0},{"conc":4096,"ok":54174,"rps":9028,"p99_us":1551080,"fail":0},{"conc":4096,"ok":55199,"rps":9199,"p99_us":608950,"fail":0},{"conc":4096,"ok":53561,"rps":8925,"p99_us":902086,"fail":0},{"conc":8192,"ok":53548,"rps":8923,"p99_us":820094,"fail":0},{"conc":8192,"ok":53060,"rps":8841,"p99_us":895983,"fail":0},{"conc":8192,"ok":52085,"rps":8680,"p99_us":1768069,"fail":0},{"conc":16384,"ok":49411,"rps":8233,"p99_us":3522568,"fail":0},{"conc":16384,"ok":50517,"rps":8418,"p99_us":2712477,"fail":0},{"conc":16384,"ok":50457,"rps":8408,"p99_us":2860866,"fail":0},{"conc":32768,"ok":19655,"rps":3269,"p99_us":903527,"fail":0},{"conc":32768,"ok":25279,"rps":4212,"p99_us":5703402,"fail":0},{"conc":32768,"ok":32784,"rps":5463,"p99_us":5485205,"fail":0}],"rungs":[{"conc":1,"rps":{"value":3395,"certified":true,"suppressed":false},"p99_us":326,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":5288,"certified":true,"suppressed":false},"p99_us":415,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":8875,"certified":true,"suppressed":false},"p99_us":606,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":9824,"certified":true,"suppressed":false},"p99_us":1214,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":9838,"certified":true,"suppressed":false},"p99_us":2557,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":9783,"certified":true,"suppressed":false},"p99_us":5302,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":9709,"certified":true,"suppressed":false},"p99_us":11281,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":9829,"certified":true,"suppressed":false},"p99_us":24831,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":9852,"certified":true,"suppressed":false},"p99_us":55818,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":9629,"certified":true,"suppressed":false},"p99_us":95520,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":9277,"certified":true,"suppressed":false},"p99_us":143955,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":9190,"certified":true,"suppressed":false},"p99_us":367277,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":9028,"certified":true,"suppressed":false},"p99_us":902086,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":8841,"certified":true,"suppressed":false},"p99_us":895983,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":8408,"certified":true,"suppressed":false},"p99_us":2860866,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":4212,"certified":true,"suppressed":false},"p99_us":5485205,"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 20084 successful gateway round trip(s) and 262977 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":20084,"direct_c1_samples":262977},"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":406,"certified":true,"suppressed":false},"added_ttft_p99_us":{"value":477,"certified":true,"suppressed":false},"added_gap_p50_us":{"value":2,"certified":true,"suppressed":false},"added_gap_p99_us":{"value":3,"certified":true,"suppressed":false},"streams_sustained_fps":{"value":25279.8938149557,"certified":true,"suppressed":false,"headroom":0.5415646457910123,"rig_ceiling":46679.36507936508},"streams_sustained":{"value":919,"certified":true,"suppressed":false,"headroom":0.5415646457910123,"rig_ceiling":46679.36507936508},"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":12.7421875,"certified":true,"suppressed":false},"steady_state_rss_mib":{"value":20.3515625,"certified":true,"suppressed":false},"recovered_rss_mib":{"value":20.21484375,"certified":true,"suppressed":false},"peak_rss_mib":{"value":21.0078125,"certified":true,"suppressed":false},"peak_rss_hwm_mib":{"value":20.5625,"certified":true,"suppressed":false},"plateaued":false,"time_to_plateau_s":{"value":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"memory was still oscillating with no net trend - it returns to where it was, which is not a leak when the 300s load window closed, so it has no time-to-plateau; its rate is published beside this"},"growth_rate_mib_per_min":{"value":-0.01099763747584857,"certified":true,"suppressed":false},"load_s":{"value":300,"certified":true,"suppressed":false},"shape":{"value":0,"certified":true,"suppressed":false},"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,"certified":true,"suppressed":false},"rss_series":[{"t_s":0,"rss_mib":12.8515625},{"t_s":2,"rss_mib":19.19140625},{"t_s":2,"rss_mib":19.1953125},{"t_s":4,"rss_mib":19.40625},{"t_s":5,"rss_mib":19.40625},{"t_s":8,"rss_mib":19.8203125},{"t_s":8,"rss_mib":19.8203125},{"t_s":11,"rss_mib":19.83203125},{"t_s":11,"rss_mib":19.84375},{"t_s":12,"rss_mib":19.88671875},{"t_s":15,"rss_mib":19.88671875},{"t_s":17,"rss_mib":19.890625},{"t_s":17,"rss_mib":19.890625},{"t_s":20,"rss_mib":20.046875},{"t_s":20,"rss_mib":20.046875},{"t_s":23,"rss_mib":20.15625},{"t_s":24,"rss_mib":20.15625},{"t_s":24,"rss_mib":20.1640625},{"t_s":26,"rss_mib":20.1640625},{"t_s":30,"rss_mib":20.1640625},{"t_s":32,"rss_mib":20.1640625},{"t_s":35,"rss_mib":20.1640625},{"t_s":39,"rss_mib":20.1640625},{"t_s":41,"rss_mib":20.1640625},{"t_s":42,"rss_mib":20.5546875},{"t_s":45,"rss_mib":20.5546875},{"t_s":48,"rss_mib":20.5546875},{"t_s":48,"rss_mib":20.33203125},{"t_s":50,"rss_mib":20.33203125},{"t_s":54,"rss_mib":20.33203125},{"t_s":54,"rss_mib":20.0625},{"t_s":56,"rss_mib":20.08203125},{"t_s":59,"rss_mib":20.0859375},{"t_s":60,"rss_mib":20.0859375},{"t_s":63,"rss_mib":20.0859375},{"t_s":65,"rss_mib":20.1015625},{"t_s":65,"rss_mib":20.1015625},{"t_s":69,"rss_mib":20.1015625},{"t_s":69,"rss_mib":20.1328125},{"t_s":72,"rss_mib":20.1328125},{"t_s":75,"rss_mib":20.1328125},{"t_s":75,"rss_mib":20.19140625},{"t_s":78,"rss_mib":20.19140625},{"t_s":80,"rss_mib":20.19140625},{"t_s":83,"rss_mib":20.21484375},{"t_s":84,"rss_mib":20.21484375},{"t_s":87,"rss_mib":20.21484375},{"t_s":89,"rss_mib":20.22265625},{"t_s":90,"rss_mib":20.22265625},{"t_s":93,"rss_mib":20.22265625},{"t_s":95,"rss_mib":20.2421875},{"t_s":96,"rss_mib":20.2421875},{"t_s":99,"rss_mib":20.2421875},{"t_s":102,"rss_mib":20.2421875},{"t_s":103,"rss_mib":20.625},{"t_s":105,"rss_mib":20.625},{"t_s":106,"rss_mib":20.62890625},{"t_s":108,"rss_mib":20.62890625},{"t_s":108,"rss_mib":20.2421875},{"t_s":111,"rss_mib":20.2421875},{"t_s":114,"rss_mib":20.2421875},{"t_s":114,"rss_mib":20.921875},{"t_s":117,"rss_mib":20.921875},{"t_s":120,"rss_mib":20.921875},{"t_s":120,"rss_mib":20.20703125},{"t_s":123,"rss_mib":20.21484375},{"t_s":124,"rss_mib":20.234375},{"t_s":126,"rss_mib":20.234375},{"t_s":126,"rss_mib":20.265625},{"t_s":129,"rss_mib":20.265625},{"t_s":132,"rss_mib":20.265625},{"t_s":135,"rss_mib":20.265625},{"t_s":138,"rss_mib":20.265625},{"t_s":141,"rss_mib":20.265625},{"t_s":144,"rss_mib":20.265625},{"t_s":145,"rss_mib":20.3046875},{"t_s":147,"rss_mib":20.3046875},{"t_s":150,"rss_mib":20.3046875},{"t_s":153,"rss_mib":20.3046875},{"t_s":156,"rss_mib":20.3046875},{"t_s":159,"rss_mib":20.3046875},{"t_s":162,"rss_mib":20.3046875},{"t_s":165,"rss_mib":20.3046875},{"t_s":168,"rss_mib":20.3046875},{"t_s":168,"rss_mib":20.66796875},{"t_s":171,"rss_mib":20.66796875},{"t_s":174,"rss_mib":20.66796875},{"t_s":174,"rss_mib":20.3828125},{"t_s":177,"rss_mib":20.3828125},{"t_s":180,"rss_mib":20.3828125},{"t_s":180,"rss_mib":21.0078125},{"t_s":183,"rss_mib":21.0078125},{"t_s":186,"rss_mib":21.0078125},{"t_s":186,"rss_mib":20.32421875},{"t_s":189,"rss_mib":20.3515625},{"t_s":189,"rss_mib":20.39453125},{"t_s":192,"rss_mib":20.39453125},{"t_s":195,"rss_mib":20.39453125},{"t_s":197,"rss_mib":20.3984375},{"t_s":198,"rss_mib":20.3984375},{"t_s":199,"rss_mib":20.7265625},{"t_s":201,"rss_mib":20.7265625},{"t_s":204,"rss_mib":20.7265625},{"t_s":207,"rss_mib":20.7265625},{"t_s":210,"rss_mib":20.7265625},{"t_s":210,"rss_mib":20.2578125},{"t_s":213,"rss_mib":20.2578125},{"t_s":216,"rss_mib":20.2578125},{"t_s":216,"rss_mib":20.34765625},{"t_s":219,"rss_mib":20.34765625},{"t_s":222,"rss_mib":20.34765625},{"t_s":225,"rss_mib":20.34765625},{"t_s":228,"rss_mib":20.34765625},{"t_s":231,"rss_mib":20.34765625},{"t_s":234,"rss_mib":20.34765625},{"t_s":234,"rss_mib":20.3515625},{"t_s":237,"rss_mib":20.3515625},{"t_s":240,"rss_mib":20.3515625},{"t_s":243,"rss_mib":20.3515625},{"t_s":246,"rss_mib":20.3515625},{"t_s":249,"rss_mib":20.3515625},{"t_s":252,"rss_mib":20.3515625},{"t_s":254,"rss_mib":20.35546875},{"t_s":255,"rss_mib":20.35546875},{"t_s":258,"rss_mib":20.35546875},{"t_s":261,"rss_mib":20.35546875},{"t_s":264,"rss_mib":20.35546875},{"t_s":264,"rss_mib":20.63671875},{"t_s":267,"rss_mib":20.63671875},{"t_s":270,"rss_mib":20.63671875},{"t_s":270,"rss_mib":20.30859375},{"t_s":273,"rss_mib":20.3125},{"t_s":276,"rss_mib":20.3125},{"t_s":279,"rss_mib":20.3125},{"t_s":282,"rss_mib":20.3125},{"t_s":284,"rss_mib":20.6328125},{"t_s":285,"rss_mib":20.6328125},{"t_s":288,"rss_mib":20.6328125},{"t_s":288,"rss_mib":20.30078125},{"t_s":291,"rss_mib":20.39453125},{"t_s":293,"rss_mib":20.47265625},{"t_s":294,"rss_mib":20.47265625},{"t_s":294,"rss_mib":20.19921875},{"t_s":297,"rss_mib":20.21484375},{"t_s":300,"rss_mib":20.21484375},{"t_s":303,"rss_mib":20.21484375},{"t_s":306,"rss_mib":20.21484375},{"t_s":309,"rss_mib":20.21484375},{"t_s":312,"rss_mib":20.21484375},{"t_s":315,"rss_mib":20.21484375},{"t_s":318,"rss_mib":20.21484375},{"t_s":321,"rss_mib":20.21484375},{"t_s":324,"rss_mib":20.21484375},{"t_s":327,"rss_mib":20.21484375},{"t_s":330,"rss_mib":20.21484375},{"t_s":333,"rss_mib":20.21484375},{"t_s":336,"rss_mib":20.21484375},{"t_s":339,"rss_mib":20.21484375},{"t_s":342,"rss_mib":20.21484375},{"t_s":345,"rss_mib":20.21484375},{"t_s":348,"rss_mib":20.21484375},{"t_s":351,"rss_mib":20.21484375},{"t_s":354,"rss_mib":20.21484375},{"t_s":357,"rss_mib":20.21484375},{"t_s":360,"rss_mib":20.21484375}],"idle_rss_series":[{"t_s":0,"rss_mib":12.7421875},{"t_s":0,"rss_mib":12.7421875},{"t_s":0,"rss_mib":12.7421875},{"t_s":1,"rss_mib":12.7421875},{"t_s":1,"rss_mib":12.7421875},{"t_s":2,"rss_mib":12.7421875},{"t_s":2,"rss_mib":12.7421875},{"t_s":3,"rss_mib":12.7421875},{"t_s":3,"rss_mib":12.7421875},{"t_s":4,"rss_mib":12.7421875},{"t_s":5,"rss_mib":12.7421875},{"t_s":5,"rss_mib":12.7421875},{"t_s":5,"rss_mib":12.7421875},{"t_s":6,"rss_mib":12.7421875},{"t_s":6,"rss_mib":12.7421875},{"t_s":7,"rss_mib":12.7421875},{"t_s":7,"rss_mib":12.7421875},{"t_s":8,"rss_mib":12.7421875},{"t_s":8,"rss_mib":12.7421875},{"t_s":9,"rss_mib":12.7421875},{"t_s":9,"rss_mib":12.7421875},{"t_s":10,"rss_mib":12.7421875},{"t_s":10,"rss_mib":12.7421875},{"t_s":11,"rss_mib":12.7421875},{"t_s":11,"rss_mib":12.7421875},{"t_s":12,"rss_mib":12.7421875},{"t_s":12,"rss_mib":12.7421875},{"t_s":13,"rss_mib":12.7421875},{"t_s":13,"rss_mib":12.7421875},{"t_s":14,"rss_mib":12.7421875},{"t_s":14,"rss_mib":12.7421875},{"t_s":15,"rss_mib":12.7421875},{"t_s":16,"rss_mib":12.7421875},{"t_s":16,"rss_mib":12.7421875},{"t_s":16,"rss_mib":12.7421875},{"t_s":17,"rss_mib":12.7421875},{"t_s":17,"rss_mib":12.7421875},{"t_s":18,"rss_mib":12.7421875},{"t_s":18,"rss_mib":12.7421875},{"t_s":19,"rss_mib":12.7421875},{"t_s":19,"rss_mib":12.7421875},{"t_s":20,"rss_mib":12.7421875},{"t_s":21,"rss_mib":12.7421875},{"t_s":21,"rss_mib":12.7421875},{"t_s":21,"rss_mib":12.7421875},{"t_s":22,"rss_mib":12.7421875},{"t_s":22,"rss_mib":12.7421875},{"t_s":23,"rss_mib":12.7421875},{"t_s":23,"rss_mib":12.7421875},{"t_s":24,"rss_mib":12.7421875},{"t_s":24,"rss_mib":12.7421875},{"t_s":25,"rss_mib":12.7421875},{"t_s":26,"rss_mib":12.7421875},{"t_s":26,"rss_mib":12.7421875},{"t_s":27,"rss_mib":12.7421875},{"t_s":27,"rss_mib":12.7421875},{"t_s":27,"rss_mib":12.7421875},{"t_s":28,"rss_mib":12.7421875},{"t_s":28,"rss_mib":12.7421875},{"t_s":29,"rss_mib":12.7421875},{"t_s":29,"rss_mib":12.7421875},{"t_s":30,"rss_mib":12.7421875},{"t_s":31,"rss_mib":12.7421875},{"t_s":31,"rss_mib":12.7421875},{"t_s":32,"rss_mib":12.7421875},{"t_s":32,"rss_mib":12.7421875},{"t_s":32,"rss_mib":12.7421875},{"t_s":33,"rss_mib":12.7421875},{"t_s":33,"rss_mib":12.7421875},{"t_s":34,"rss_mib":12.7421875},{"t_s":34,"rss_mib":12.7421875},{"t_s":35,"rss_mib":12.7421875},{"t_s":36,"rss_mib":12.7421875},{"t_s":36,"rss_mib":12.7421875},{"t_s":37,"rss_mib":12.7421875},{"t_s":37,"rss_mib":12.7421875},{"t_s":38,"rss_mib":12.7421875},{"t_s":38,"rss_mib":12.7421875},{"t_s":38,"rss_mib":12.7421875},{"t_s":39,"rss_mib":12.7421875},{"t_s":39,"rss_mib":12.7421875},{"t_s":40,"rss_mib":12.7421875},{"t_s":41,"rss_mib":12.7421875},{"t_s":41,"rss_mib":12.7421875},{"t_s":42,"rss_mib":12.7421875},{"t_s":42,"rss_mib":12.7421875},{"t_s":43,"rss_mib":12.7421875},{"t_s":43,"rss_mib":12.7421875},{"t_s":43,"rss_mib":12.7421875},{"t_s":44,"rss_mib":12.7421875},{"t_s":44,"rss_mib":12.7421875},{"t_s":45,"rss_mib":12.7421875},{"t_s":45,"rss_mib":12.7421875},{"t_s":46,"rss_mib":12.7421875},{"t_s":47,"rss_mib":12.7421875},{"t_s":47,"rss_mib":12.7421875},{"t_s":48,"rss_mib":12.7421875},{"t_s":48,"rss_mib":12.7421875},{"t_s":49,"rss_mib":12.7421875},{"t_s":49,"rss_mib":12.7421875},{"t_s":49,"rss_mib":12.7421875},{"t_s":50,"rss_mib":12.7421875},{"t_s":50,"rss_mib":12.7421875},{"t_s":51,"rss_mib":12.7421875},{"t_s":52,"rss_mib":12.7421875},{"t_s":52,"rss_mib":12.7421875},{"t_s":53,"rss_mib":12.7421875},{"t_s":53,"rss_mib":12.7421875},{"t_s":54,"rss_mib":12.7421875},{"t_s":54,"rss_mib":12.7421875},{"t_s":54,"rss_mib":12.7421875},{"t_s":55,"rss_mib":12.7421875},{"t_s":55,"rss_mib":12.7421875},{"t_s":56,"rss_mib":12.7421875},{"t_s":57,"rss_mib":12.7421875},{"t_s":57,"rss_mib":12.7421875},{"t_s":58,"rss_mib":12.7421875},{"t_s":58,"rss_mib":12.7421875},{"t_s":59,"rss_mib":12.7421875},{"t_s":59,"rss_mib":12.7421875},{"t_s":59,"rss_mib":12.7421875}],"idle_window_s":60,"recovery_window_s":30},"timings_s":{"added_latency":12.008841495,"cost":6.02603779,"memory":422.730140067,"streaming":2.635493549,"streams_sustained":157.775763296,"throughput":583.542302027},"upstream_transport":"http","absences":{"memory.idle_shape":{"reason":"not_measured","detail":"a settled idle window has no unsettled shape to describe"},"memory.time_to_plateau_s":{"reason":"not_measured","detail":"memory was still oscillating with no net trend - it returns to where it was, which is not a leak when the 300s load window closed, so it has no time-to-plateau; its rate is published beside this"}}},"openai-responses":{"served":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/responses","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":false,"served":false,"egress_config":null,"serve_error":"","cells":{"anthropic":{"served":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/messages","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v2/chat","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1beta/models/bench-openai:generateContent","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/model/bench-openai/converse","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/responses","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/chat/completions","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","body_snippet":"","probe_note":null,"perf_dropped":null,"perf":null,"stream":null,"memory":null,"timings_s":null,"upstream_transport":null,"absences":{}}}},"cohere":{"configurable":false,"served":false,"egress_config":null,"serve_error":"","cells":{"openai-responses":{"served":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/responses","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/chat/completions","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v2/chat","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1beta/models/bench-openai:generateContent","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/v1/messages","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","status":"","path":"/model/bench-openai/converse","verdict_note":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"outcome":"pass"},"somaxconn":65535,"tcp_max_syn_backlog":65535},"measured_at":"2026-11-10T17:39:51Z","started_at":null,"finished_at":null,"duration_s":null,"phase_s":null,"build_env":null},"matrix_from_snapshot":true,"snapshot_file":"result_litellm-rust_2026-11-10T17-39-51Z.json","ootb_config":"general_settings:\n  master_key: \"gwbench\"\nmodel_list:\n  - model_name: bench-openai\n    litellm_params:\n      model: openai_like/gpt-4o-mini\n      api_base: http://127.0.0.1:8000/v1\n      api_key: dummy\n  - model_name: bench-responses\n    litellm_params:\n      model: openai/gpt-4o-mini\n      api_base: http://127.0.0.1:8000/v1\n      api_key: dummy\n  - model_name: bench-anthropic\n    litellm_params:\n      model: anthropic/claude-3-5-sonnet-20241022\n      api_base: http://127.0.0.1:8000\n      api_key: dummy\n  - model_name: bench-bedrock\n    litellm_params:\n      model: bedrock/amazon.nova-micro-v1:0\n      api_base: http://127.0.0.1:8000\n      api_key: dummy\n      aws_region_name: us-east-1\n","best_cell":{"path":{"ingress":"openai","egress":"openai","dialect":"openai"},"source":{"kind":"matrix","sweep":"6x6-diagonal","build":"src/litellm-rust/target/release/litellm-gateway","measured_at":"2026-11-10T17:39:51Z"},"added_latency_p50_us":{"value":249,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":279,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":306,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":27,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":272.6013913681787,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":3668.359853121175,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.6790579018170796,"certified":true,"suppressed":false},"cost_window_ok":{"value":59941,"certified":true,"suppressed":false},"cost_window_rps":{"value":9989,"certified":true,"suppressed":false},"cost_threads":{"value":5,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":8404,"certified":true,"suppressed":false},"concurrency":4,"p99_us":588,"first_disqualified_conc":8,"lower_bound":false,"samples":50435},{"bound_ms":5,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":32,"lower_bound":false,"samples":60306},{"bound_ms":10,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":64,"lower_bound":false,"samples":60306},{"bound_ms":50,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":256,"lower_bound":false,"samples":60306},{"bound_ms":100,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":1024,"lower_bound":false,"samples":60306},{"bound_ms":null,"rps":{"value":10049,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2486,"first_disqualified_conc":null,"lower_bound":false,"samples":60306}],"sweep":[{"conc":1,"ok":22499,"rps":3748,"p99_us":301,"fail":0},{"conc":1,"ok":22443,"rps":3738,"p99_us":300,"fail":0},{"conc":1,"ok":22367,"rps":3726,"p99_us":302,"fail":0},{"conc":2,"ok":32753,"rps":5458,"p99_us":410,"fail":0},{"conc":2,"ok":32736,"rps":5455,"p99_us":409,"fail":0},{"conc":2,"ok":32631,"rps":5437,"p99_us":409,"fail":0},{"conc":4,"ok":50172,"rps":8361,"p99_us":589,"fail":0},{"conc":4,"ok":50353,"rps":8391,"p99_us":593,"fail":0},{"conc":4,"ok":50435,"rps":8404,"p99_us":588,"fail":0},{"conc":8,"ok":60135,"rps":10021,"p99_us":1218,"fail":0},{"conc":8,"ok":60158,"rps":10025,"p99_us":1223,"fail":0},{"conc":8,"ok":60188,"rps":10030,"p99_us":1221,"fail":0},{"conc":16,"ok":60306,"rps":10049,"p99_us":2486,"fail":0},{"conc":16,"ok":60202,"rps":10032,"p99_us":2503,"fail":0},{"conc":16,"ok":60205,"rps":10033,"p99_us":2499,"fail":0},{"conc":32,"ok":59745,"rps":9956,"p99_us":5251,"fail":0},{"conc":32,"ok":59963,"rps":9992,"p99_us":5176,"fail":0},{"conc":32,"ok":59839,"rps":9972,"p99_us":5236,"fail":0},{"conc":64,"ok":59403,"rps":9900,"p99_us":11251,"fail":0},{"conc":64,"ok":59179,"rps":9862,"p99_us":11128,"fail":0},{"conc":64,"ok":59226,"rps":9870,"p99_us":11116,"fail":0},{"conc":128,"ok":59428,"rps":9903,"p99_us":25558,"fail":0},{"conc":128,"ok":59277,"rps":9878,"p99_us":25080,"fail":0},{"conc":128,"ok":59326,"rps":9886,"p99_us":25057,"fail":0},{"conc":256,"ok":59141,"rps":9855,"p99_us":54342,"fail":0},{"conc":256,"ok":59008,"rps":9833,"p99_us":60650,"fail":0},{"conc":256,"ok":58946,"rps":9823,"p99_us":57205,"fail":0},{"conc":512,"ok":57998,"rps":9664,"p99_us":94929,"fail":0},{"conc":512,"ok":58039,"rps":9672,"p99_us":94698,"fail":0},{"conc":512,"ok":58127,"rps":9687,"p99_us":95092,"fail":0},{"conc":1024,"ok":56975,"rps":9494,"p99_us":146358,"fail":0},{"conc":1024,"ok":57324,"rps":9552,"p99_us":151112,"fail":0},{"conc":1024,"ok":57310,"rps":9549,"p99_us":147598,"fail":0},{"conc":2048,"ok":56728,"rps":9453,"p99_us":412827,"fail":0},{"conc":2048,"ok":56479,"rps":9411,"p99_us":344184,"fail":0},{"conc":2048,"ok":56128,"rps":9353,"p99_us":343443,"fail":0},{"conc":4096,"ok":55274,"rps":9211,"p99_us":824938,"fail":0},{"conc":4096,"ok":56042,"rps":9339,"p99_us":898347,"fail":0},{"conc":4096,"ok":55560,"rps":9259,"p99_us":669703,"fail":0},{"conc":8192,"ok":54530,"rps":9087,"p99_us":1884573,"fail":0},{"conc":8192,"ok":54911,"rps":9151,"p99_us":1996218,"fail":0},{"conc":8192,"ok":54551,"rps":9091,"p99_us":2235636,"fail":0},{"conc":16384,"ok":53775,"rps":8961,"p99_us":3860409,"fail":0},{"conc":16384,"ok":53451,"rps":8908,"p99_us":2392444,"fail":0},{"conc":16384,"ok":52935,"rps":8821,"p99_us":2261573,"fail":0},{"conc":32768,"ok":31085,"rps":5180,"p99_us":5817043,"fail":0},{"conc":32768,"ok":16399,"rps":2730,"p99_us":1118348,"fail":0},{"conc":32768,"ok":38059,"rps":6342,"p99_us":5856115,"fail":0}],"rungs":[{"conc":1,"rps":{"value":3738,"certified":true,"suppressed":false},"p99_us":301,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":5455,"certified":true,"suppressed":false},"p99_us":409,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":8391,"certified":true,"suppressed":false},"p99_us":589,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":10025,"certified":true,"suppressed":false},"p99_us":1221,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":10033,"certified":true,"suppressed":false},"p99_us":2499,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":9972,"certified":true,"suppressed":false},"p99_us":5236,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":9870,"certified":true,"suppressed":false},"p99_us":11128,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":9886,"certified":true,"suppressed":false},"p99_us":25080,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":9833,"certified":true,"suppressed":false},"p99_us":57205,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":9672,"certified":true,"suppressed":false},"p99_us":94929,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":9549,"certified":true,"suppressed":false},"p99_us":147598,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":9411,"certified":true,"suppressed":false},"p99_us":344184,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":9259,"certified":true,"suppressed":false},"p99_us":824938,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":9091,"certified":true,"suppressed":false},"p99_us":1996218,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":8908,"certified":true,"suppressed":false},"p99_us":2392444,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":5180,"certified":true,"suppressed":false},"p99_us":5817043,"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 22021 successful gateway round trip(s) and 267021 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":22021,"direct_c1_samples":267021,"rps_per_dollar":{"value":61574.75490196078,"certified":true,"suppressed":false},"cost_per_million_usd":{"value":0.004511228314591834,"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":"src/litellm-rust/target/release/litellm-gateway","measured_at":"2026-11-10T17:39:51Z"},"added_latency_p50_us":{"value":253,"certified":true,"suppressed":false},"added_latency_p99_us":{"value":283,"certified":true,"suppressed":false},"gateway_c1_p99_us":{"value":311,"certified":true,"suppressed":false},"direct_c1_p99_us":{"value":28,"certified":true,"suppressed":false},"cpu_us_per_request":{"value":277.0705901801951,"certified":true,"suppressed":false},"rps_per_cpu_second":{"value":3609.1885441527443,"certified":true,"suppressed":false},"cost_window_conc":{"value":8,"certified":true,"suppressed":false},"cost_core_utilisation":{"value":0.6965431268306815,"certified":true,"suppressed":false},"cost_window_ok":{"value":60490,"certified":true,"suppressed":false},"cost_window_rps":{"value":10080,"certified":true,"suppressed":false},"cost_threads":{"value":5,"certified":true,"suppressed":false},"cost_nonvol_ctxt_per_request":{"value":0,"certified":true,"suppressed":false},"cost_majflt":{"value":0,"certified":true,"suppressed":false},"frontier":[{"bound_ms":1,"rps":{"value":8572,"certified":true,"suppressed":false},"concurrency":4,"p99_us":594,"first_disqualified_conc":8,"lower_bound":false,"samples":51441},{"bound_ms":5,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":32,"lower_bound":false,"samples":60500},{"bound_ms":10,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":64,"lower_bound":false,"samples":60500},{"bound_ms":50,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":256,"lower_bound":false,"samples":60500},{"bound_ms":100,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":1024,"lower_bound":false,"samples":60500},{"bound_ms":null,"rps":{"value":10082,"certified":true,"suppressed":false},"concurrency":16,"p99_us":2482,"first_disqualified_conc":null,"lower_bound":false,"samples":60500}],"sweep":[{"conc":1,"ok":21640,"rps":3606,"p99_us":311,"fail":0},{"conc":1,"ok":21650,"rps":3607,"p99_us":311,"fail":0},{"conc":1,"ok":21569,"rps":3594,"p99_us":312,"fail":0},{"conc":2,"ok":31855,"rps":5308,"p99_us":418,"fail":0},{"conc":2,"ok":31701,"rps":5282,"p99_us":419,"fail":0},{"conc":2,"ok":31644,"rps":5273,"p99_us":420,"fail":0},{"conc":4,"ok":51006,"rps":8499,"p99_us":595,"fail":0},{"conc":4,"ok":51441,"rps":8572,"p99_us":594,"fail":0},{"conc":4,"ok":51316,"rps":8551,"p99_us":593,"fail":0},{"conc":8,"ok":60286,"rps":10046,"p99_us":1203,"fail":0},{"conc":8,"ok":60318,"rps":10051,"p99_us":1209,"fail":0},{"conc":8,"ok":60327,"rps":10053,"p99_us":1209,"fail":0},{"conc":16,"ok":60324,"rps":10052,"p99_us":2493,"fail":0},{"conc":16,"ok":60500,"rps":10082,"p99_us":2482,"fail":0},{"conc":16,"ok":60325,"rps":10053,"p99_us":2496,"fail":0},{"conc":32,"ok":59850,"rps":9974,"p99_us":5218,"fail":0},{"conc":32,"ok":59868,"rps":9977,"p99_us":5187,"fail":0},{"conc":32,"ok":59897,"rps":9982,"p99_us":5199,"fail":0},{"conc":64,"ok":59227,"rps":9870,"p99_us":11146,"fail":0},{"conc":64,"ok":59387,"rps":9897,"p99_us":11181,"fail":0},{"conc":64,"ok":59404,"rps":9900,"p99_us":11163,"fail":0},{"conc":128,"ok":59602,"rps":9932,"p99_us":25187,"fail":0},{"conc":128,"ok":59509,"rps":9916,"p99_us":25168,"fail":0},{"conc":128,"ok":59427,"rps":9904,"p99_us":24757,"fail":0},{"conc":256,"ok":59085,"rps":9846,"p99_us":55035,"fail":0},{"conc":256,"ok":59251,"rps":9874,"p99_us":57249,"fail":0},{"conc":256,"ok":59368,"rps":9893,"p99_us":57642,"fail":0},{"conc":512,"ok":58011,"rps":9668,"p99_us":94197,"fail":0},{"conc":512,"ok":58049,"rps":9674,"p99_us":93858,"fail":0},{"conc":512,"ok":58273,"rps":9711,"p99_us":94318,"fail":0},{"conc":1024,"ok":56680,"rps":9445,"p99_us":138751,"fail":0},{"conc":1024,"ok":57008,"rps":9499,"p99_us":137676,"fail":0},{"conc":1024,"ok":56798,"rps":9464,"p99_us":132484,"fail":0},{"conc":2048,"ok":55560,"rps":9258,"p99_us":328520,"fail":0},{"conc":2048,"ok":55986,"rps":9329,"p99_us":367636,"fail":0},{"conc":2048,"ok":56137,"rps":9354,"p99_us":340707,"fail":0},{"conc":4096,"ok":55199,"rps":9199,"p99_us":854765,"fail":0},{"conc":4096,"ok":56235,"rps":9372,"p99_us":465530,"fail":0},{"conc":4096,"ok":54901,"rps":9148,"p99_us":1368193,"fail":0},{"conc":8192,"ok":54255,"rps":9041,"p99_us":2169098,"fail":0},{"conc":8192,"ok":53775,"rps":8961,"p99_us":2316246,"fail":0},{"conc":8192,"ok":54267,"rps":9043,"p99_us":1786177,"fail":0},{"conc":16384,"ok":52209,"rps":8700,"p99_us":2878906,"fail":0},{"conc":16384,"ok":52949,"rps":8824,"p99_us":2916402,"fail":0},{"conc":16384,"ok":52168,"rps":8693,"p99_us":2634896,"fail":0},{"conc":32768,"ok":16730,"rps":2780,"p99_us":1164793,"fail":0},{"conc":32768,"ok":32916,"rps":5485,"p99_us":5439856,"fail":0},{"conc":32768,"ok":32440,"rps":5405,"p99_us":3847695,"fail":0}],"rungs":[{"conc":1,"rps":{"value":3606,"certified":true,"suppressed":false},"p99_us":311,"windows":3,"clean_windows":3},{"conc":2,"rps":{"value":5282,"certified":true,"suppressed":false},"p99_us":419,"windows":3,"clean_windows":3},{"conc":4,"rps":{"value":8551,"certified":true,"suppressed":false},"p99_us":594,"windows":3,"clean_windows":3},{"conc":8,"rps":{"value":10051,"certified":true,"suppressed":false},"p99_us":1209,"windows":3,"clean_windows":3},{"conc":16,"rps":{"value":10053,"certified":true,"suppressed":false},"p99_us":2493,"windows":3,"clean_windows":3},{"conc":32,"rps":{"value":9977,"certified":true,"suppressed":false},"p99_us":5199,"windows":3,"clean_windows":3},{"conc":64,"rps":{"value":9897,"certified":true,"suppressed":false},"p99_us":11163,"windows":3,"clean_windows":3},{"conc":128,"rps":{"value":9916,"certified":true,"suppressed":false},"p99_us":25168,"windows":3,"clean_windows":3},{"conc":256,"rps":{"value":9874,"certified":true,"suppressed":false},"p99_us":57249,"windows":3,"clean_windows":3},{"conc":512,"rps":{"value":9674,"certified":true,"suppressed":false},"p99_us":94197,"windows":3,"clean_windows":3},{"conc":1024,"rps":{"value":9464,"certified":true,"suppressed":false},"p99_us":137676,"windows":3,"clean_windows":3},{"conc":2048,"rps":{"value":9329,"certified":true,"suppressed":false},"p99_us":340707,"windows":3,"clean_windows":3},{"conc":4096,"rps":{"value":9199,"certified":true,"suppressed":false},"p99_us":854765,"windows":3,"clean_windows":3},{"conc":8192,"rps":{"value":9041,"certified":true,"suppressed":false},"p99_us":2169098,"windows":3,"clean_windows":3},{"conc":16384,"rps":{"value":8700,"certified":true,"suppressed":false},"p99_us":2878906,"windows":3,"clean_windows":3},{"conc":32768,"rps":{"value":5405,"certified":true,"suppressed":false},"p99_us":3847695,"windows":3,"clean_windows":3}],"egress_reverified":true,"c1_note":"the c=1 percentiles are taken over 21651 successful gateway round trip(s) and 262942 direct-to-mock round trip(s), each leg a single clean window with no failures","gateway_c1_samples":21651,"direct_c1_samples":262942},"measured_at":"2026-11-10T17:39:51.000Z","lane_measured_at":{"perf":"2026-11-10T17:39:51Z","xlate":"2026-11-10T17:39:51Z","memory":"2026-11-10T17:39:51Z"},"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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"outcome":"pass"},"somaxconn":65535,"tcp_max_syn_backlog":65535},"contributed_by":[],"stale":false,"engine":{"sha":"d240badd966933fa19eec0606b2d1f3945f18ad2","short":"d240bad","current":true},"history_family":"litellm-rust","history_trends":{}},"current_metadata":{"stars":63241,"stars_as_of":"2026-11-14","first_commit":"2023-07-27"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:anthropic:bedrock","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:cohere:bedrock","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:bedrock:bedrock","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:gemini:bedrock","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai:bedrock","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_served","raw_state":"failed","reason":"failed","detail":"probed and answered failed (HTTP 501)","evidence_scope":"Recorded run evidence","probe_http_status":"501","response_excerpt":"{\"error\":{\"message\":\"unsupported by the rust path: unrecognized request parameter\",\"type\":\"api_error\",\"param\":null,\"code\":501}}"},"test":{"ingress_protocol":"openai","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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai-responses:bedrock","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:bedrock:openai","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai:openai","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{"added_latency_p50_us":{"value":249,"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":279,"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":306,"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":272.6013913681787,"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":3668.359853121175,"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":0.6790579018170796,"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":59941,"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":9989,"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":5,"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,"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":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","state":"not_measured","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":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","state":"not_measured","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":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","state":"not_measured","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":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","state":"not_measured","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":0,"certified":true,"suppressed":false,"note":"measured_failure","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":0,"certified":true,"suppressed":false,"note":"measured_failure","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":12.73828125,"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":18.9921875,"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."},"recovered_rss_mib":{"value":18.99609375,"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":19.875,"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."},"peak_rss_hwm_mib":{"value":19.484375,"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":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"memory was still oscillating with no net trend - it returns to where it was, which is not a leak when the 300s load window closed, so it has no time-to-plateau; its rate is published beside this","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."},"growth_rate_mib_per_min":{"value":-0.1791634760562065,"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":0,"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_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,"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":8404,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":4,"p99_us":588,"first_disqualified_conc":8,"lower_bound":false,"samples":50435},{"bound_ms":5,"rps":{"value":10049,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2486,"first_disqualified_conc":32,"lower_bound":false,"samples":60306},{"bound_ms":10,"rps":{"value":10049,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2486,"first_disqualified_conc":64,"lower_bound":false,"samples":60306},{"bound_ms":50,"rps":{"value":10049,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2486,"first_disqualified_conc":256,"lower_bound":false,"samples":60306},{"bound_ms":100,"rps":{"value":10049,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2486,"first_disqualified_conc":1024,"lower_bound":false,"samples":60306},{"bound_ms":null,"rps":{"value":10049,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2486,"first_disqualified_conc":null,"lower_bound":false,"samples":60306}],"sample_counts":{"gateway_c1_samples":22021,"direct_c1_samples":267021},"stream_rung_diagnostics":[{"conc":1,"http_status_counts":{"501":1},"stream_errors_no_response":0}],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:anthropic:openai","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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","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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai-responses:openai","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:gemini:openai","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:cohere:openai","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:gemini:openai-responses","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:cohere:openai-responses","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:bedrock:openai-responses","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai-responses:openai-responses","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{"added_latency_p50_us":{"value":267,"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":298,"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":325,"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":287.74399582885263,"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":3475.3114382785957,"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":0.7339219155366118,"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":61374,"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":10227,"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":5,"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,"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":12.7421875,"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."},"steady_state_rss_mib":{"value":19.47265625,"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":19.28515625,"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":19.5,"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":19.5,"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":0.028056430633918075,"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,"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":8050,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":4,"p99_us":705,"first_disqualified_conc":8,"lower_bound":false,"samples":48309},{"bound_ms":5,"rps":{"value":10191,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2455,"first_disqualified_conc":32,"lower_bound":false,"samples":61157},{"bound_ms":10,"rps":{"value":10191,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2455,"first_disqualified_conc":64,"lower_bound":false,"samples":61157},{"bound_ms":50,"rps":{"value":10191,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2455,"first_disqualified_conc":256,"lower_bound":false,"samples":61157},{"bound_ms":100,"rps":{"value":10191,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2455,"first_disqualified_conc":1024,"lower_bound":false,"samples":61157},{"bound_ms":null,"rps":{"value":10191,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2455,"first_disqualified_conc":null,"lower_bound":false,"samples":61157}],"sample_counts":{"gateway_c1_samples":20661,"direct_c1_samples":267569},"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai:openai-responses","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:anthropic:openai-responses","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai:anthropic","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{"added_latency_p50_us":{"value":253,"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":283,"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":311,"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":277.0705901801951,"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":3609.1885441527443,"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":0.6965431268306815,"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":60490,"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":10080,"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":5,"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,"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":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","state":"not_measured","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":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","state":"not_measured","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":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","state":"not_measured","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":"not_measured","detail":"no stream frame arrived from the gateway: it answered HTTP 501, not 2xx, so what it sent is an error response rather than a stream to time","state":"not_measured","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":0,"certified":true,"suppressed":false,"note":"measured_failure","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":0,"certified":true,"suppressed":false,"note":"measured_failure","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":12.7578125,"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":19.046875,"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":18.984375,"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":19.66015625,"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":19.33203125,"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":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"memory was still oscillating with no net trend - it returns to where it was, which is not a leak when the 300s load window closed, so it has no time-to-plateau; its rate is published beside this","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."},"growth_rate_mib_per_min":{"value":0.3275657622729096,"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":0,"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_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,"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":8572,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":4,"p99_us":594,"first_disqualified_conc":8,"lower_bound":false,"samples":51441},{"bound_ms":5,"rps":{"value":10082,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2482,"first_disqualified_conc":32,"lower_bound":false,"samples":60500},{"bound_ms":10,"rps":{"value":10082,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2482,"first_disqualified_conc":64,"lower_bound":false,"samples":60500},{"bound_ms":50,"rps":{"value":10082,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2482,"first_disqualified_conc":256,"lower_bound":false,"samples":60500},{"bound_ms":100,"rps":{"value":10082,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2482,"first_disqualified_conc":1024,"lower_bound":false,"samples":60500},{"bound_ms":null,"rps":{"value":10082,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2482,"first_disqualified_conc":null,"lower_bound":false,"samples":60500}],"sample_counts":{"gateway_c1_samples":21651,"direct_c1_samples":262942},"stream_rung_diagnostics":[{"conc":1,"http_status_counts":{"501":1},"stream_errors_no_response":0}],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:cohere:anthropic","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:bedrock:anthropic","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:gemini:anthropic","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:anthropic:anthropic","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{"added_latency_p50_us":{"value":275,"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":304,"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":332,"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":297.0902986640554,"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":3365.97998822837,"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":0.7061370243602647,"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":57188,"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":9530,"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":5,"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,"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":406,"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":477,"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":3,"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":25279.8938149557,"certified":true,"suppressed":false,"headroom":0.5415646457910123,"rig_ceiling":46679.36507936508,"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":919,"certified":true,"suppressed":false,"headroom":0.5415646457910123,"rig_ceiling":46679.36507936508,"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":12.7421875,"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."},"steady_state_rss_mib":{"value":20.3515625,"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":20.21484375,"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":21.0078125,"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":20.5625,"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":null,"certified":false,"suppressed":false,"reason":"not_measured","detail":"memory was still oscillating with no net trend - it returns to where it was, which is not a leak when the 300s load window closed, so it has no time-to-plateau; its rate is published beside this","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."},"growth_rate_mib_per_min":{"value":-0.01099763747584857,"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":0,"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_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,"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":8878,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":4,"p99_us":606,"first_disqualified_conc":8,"lower_bound":false,"samples":53276},{"bound_ms":5,"rps":{"value":9865,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2555,"first_disqualified_conc":32,"lower_bound":false,"samples":59196},{"bound_ms":10,"rps":{"value":9865,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2555,"first_disqualified_conc":64,"lower_bound":false,"samples":59196},{"bound_ms":50,"rps":{"value":9865,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2555,"first_disqualified_conc":256,"lower_bound":false,"samples":59196},{"bound_ms":100,"rps":{"value":9865,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2555,"first_disqualified_conc":1024,"lower_bound":false,"samples":59196},{"bound_ms":null,"rps":{"value":9865,"certified":true,"suppressed":false,"state":"valid","reason":null},"concurrency":16,"p99_us":2555,"first_disqualified_conc":null,"lower_bound":false,"samples":59196}],"sample_counts":{"gateway_c1_samples":20084,"direct_c1_samples":262977,"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":768,"http_status_counts":{"200":768},"stream_errors_no_response":0},{"conc":896,"http_status_counts":{"200":896},"stream_errors_no_response":0},{"conc":960,"http_status_counts":{"200":960},"stream_errors_no_response":0},{"conc":928,"http_status_counts":{"200":928},"stream_errors_no_response":0},{"conc":912,"http_status_counts":{"200":912},"stream_errors_no_response":0},{"conc":920,"http_status_counts":{"200":920},"stream_errors_no_response":0},{"conc":916,"http_status_counts":{"200":916},"stream_errors_no_response":0},{"conc":918,"http_status_counts":{"200":918},"stream_errors_no_response":0},{"conc":919,"http_status_counts":{"200":919},"stream_errors_no_response":0},{"conc":919,"http_status_counts":{"200":919},"stream_errors_no_response":0},{"conc":919,"http_status_counts":{"200":919},"stream_errors_no_response":0}],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai-responses:anthropic","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:anthropic:gemini","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:cohere:gemini","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:gemini:gemini","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:bedrock:gemini","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai-responses:gemini","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai:gemini","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai-responses:cohere","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:openai:cohere","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:cohere:cohere","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:gemini:cohere","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:anthropic:cohere","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":"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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4:bedrock:cohere","gateway_id":"litellm-rust","gateway_build":"src/litellm-rust/target/release/litellm-gateway","run_id":"run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","measured_at":"2026-11-10T17:39:51Z","status":{"state":"not_declared","raw_state":"not_configurable","reason":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-preflight.json","detail":"Pinned v1.108.3 standalone Rust source b7d579fab2818968322ef96e5b8e98dc8178fafc: gateway-inference/src/lib.rs registers Chat, Responses and Messages ingress. inference-chat/src/common_utils.rs registers OpenAI-like, Anthropic and Bedrock Converse adapters; inference-responses/src/prepare.rs supports native OpenAI Responses; inference-messages/src/common_utils.rs has native Anthropic, but no OpenAI cross-conversion. Its Bedrock Messages path uses Invoke, not this matrix's Converse wire. Native Gemini, Cohere and Bedrock ingress are not registered. OpenAI Chat to Bedrock Converse is configured and probed with the standard workload request, which includes max_tokens; the released gateway answers HTTP 501 (\"unsupported by the rust path: unrecognized request parameter\"), which the engine records as not configured. Pairs outside the routes above are not configured. Evidence: results/diagnostics/refresh-configuration/litellm-rust-exploratory-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":-13.363964282372997,"observed_absent_detail":null,"observed_absent_reason":null,"observed_rps":83343,"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,"gateway_provenance":{"binary_name":"litellm-gateway","binary_sha256":"ae1fec119755a8a2ae5b6a9d56225c34205ce7a54548cb11696b03b2557ce01b","cargo_lock_sha256":"ae9550da9f00c3ffb10ac2974d7d33c3784902c702c74c6aa85f1b83f1efcb2d","compiler":"rustc 1.99.1 (58cf8f9c3 2026-10-13)\nbinary: rustc\ncommit-hash: 58cf8f9c35b355199727627d07ae34a4e2846d67\ncommit-date: 2026-10-13\nhost: aarch64-unknown-linux-gnu\nrelease: 1.99.1\nLLVM version: 23.1.1","source_commit":"b7d579fab2818968322ef96e5b8e98dc8178fafc","source_dirty":false}},"metrics":{},"frontier":[],"provenance":{"source_url":"https://onthebench.ai/gateways/litellm-rust","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json","run_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","methodology_url":"https://onthebench.ai/gateways/method","raw_sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4"},"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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","sha256":"69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","gateway_id":"litellm-rust","measured_at":"2026-11-10T17:39:51Z","build":"src/litellm-rust/target/release/litellm-gateway","gateway_family":"litellm-rust","engine":"d240badd966933fa19eec0606b2d1f3945f18ad2","projector":"historical-matrix-v1-e1440c114a2bba3bde4a01bfa138bc6160e327576b1d93bd2dced418d56805ae","configuration_sha256":"d2eb614b13e66fffc96694c51d4fc0e95b751262894bdcbe6c335517960106d9","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-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/measurements","source_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-69d8e6ed325740d9da7f0aaf31b815418c4801eb6549b2e2c68f768a92af3ad4/raw.json"},{"id":"run-litellm-rust-50a1b780253a9655d548618bbcdd0969dd22b4f1a867714256048d1d68820eb9","sha256":"50a1b780253a9655d548618bbcdd0969dd22b4f1a867714256048d1d68820eb9","gateway_id":"litellm-rust","measured_at":"2026-08-03T21:30:01Z","build":"litellm-ai-gateway","gateway_family":"litellm-rust","engine":"f3adbb1315b26129f5e317af5279decefb1cea8f","projector":"historical-matrix-v1-e1440c114a2bba3bde4a01bfa138bc6160e327576b1d93bd2dced418d56805ae","configuration_sha256":"1ef52ce3a7b0164ce21afdf4a79a78f9c95ad2c8b5c5171fee35e8bc3ace831e","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-litellm-rust-50a1b780253a9655d548618bbcdd0969dd22b4f1a867714256048d1d68820eb9/measurements","source_url":"https://onthebench.ai/gateways/runs/run-litellm-rust-50a1b780253a9655d548618bbcdd0969dd22b4f1a867714256048d1d68820eb9","raw_result_url":"https://onthebench.ai/api/v1/runs/run-litellm-rust-50a1b780253a9655d548618bbcdd0969dd22b4f1a867714256048d1d68820eb9/raw.json"}],"generated_at":"2026-11-14T11:09:24.858Z","disclosure":{"operator":"Busbar team","operator_is_entrant":true}}
