Introduced by #1211 (b537a5344) and found immediately after it merged.
The defect
docs/USAGE.md:5186 states "They also ran a different binary. Every figure below comes from benchmarks/expert_stream_device_w0e.cpp", the C ABI harness, rather than from the vllm-server command the section publishes.
That is true of the W0e and W0f figures and false for one row. The 66.7 s/token streaming-off measurement of 16 August 2026 was itself a server run. .agents/specs/expert-streaming.md:837 records the first run as --max-model-len 512, VT_GGUF_PREFAULT=0, server entry point, and :905 records the 16 August run as the same vehicle and configuration with streaming off.
So the one figure the section presents as measured on a different binary from the reader's is the one figure measured on the same binary the reader is told to run. A reader reasoning about how much of the gap is harness-versus-server overhead gets it backwards for that row.
How it got in
The correction was authored while #1211 was still open, and was not pushed before that pull request merged. The merge outran it. That is a coordination failure rather than a review miss: the fresh review of #1211 flagged the missing binary disclosure, the repair added it, and the scoping refinement to that repair arrived a few minutes late.
Fix
Scope the sentence to the runs it is true of and name the exception, in docs/USAGE.md and in the spec paragraph that mirrors it. No figure changes and no new claim is added.
Introduced by #1211 (
b537a5344) and found immediately after it merged.The defect
docs/USAGE.md:5186states "They also ran a different binary. Every figure below comes frombenchmarks/expert_stream_device_w0e.cpp", the C ABI harness, rather than from thevllm-servercommand the section publishes.That is true of the W0e and W0f figures and false for one row. The 66.7 s/token streaming-off measurement of 16 August 2026 was itself a server run.
.agents/specs/expert-streaming.md:837records the first run as--max-model-len 512,VT_GGUF_PREFAULT=0, server entry point, and:905records the 16 August run as the same vehicle and configuration with streaming off.So the one figure the section presents as measured on a different binary from the reader's is the one figure measured on the same binary the reader is told to run. A reader reasoning about how much of the gap is harness-versus-server overhead gets it backwards for that row.
How it got in
The correction was authored while #1211 was still open, and was not pushed before that pull request merged. The merge outran it. That is a coordination failure rather than a review miss: the fresh review of #1211 flagged the missing binary disclosure, the repair added it, and the scoping refinement to that repair arrived a few minutes late.
Fix
Scope the sentence to the runs it is true of and name the exception, in
docs/USAGE.mdand in the spec paragraph that mirrors it. No figure changes and no new claim is added.