Found while rebuilding the Qwen3.8 reproduction section (#1211) against the measurements that landed in #1414 and #1427. It belongs to the ENG-EXPERT-STREAM row (#912), not to ENG-EXPERT-STREAM-DEVICE, which is why neither PR fixed it.
The contradiction
docs/BENCHMARKS.md line 8, the ENG-EXPERT-STREAM (#912) row, says the streaming-ON decode figure is VOID — the step clock had no caller, so the cache was never exercised as measured.
docs/BENCHMARKS.md line 9, the ENG-EXPERT-STREAM-DEVICE (#1124, #1299) row, records 11.05 s/token at 4000 slots, live cache, decode-phase exhausted delta 0, explicitly as replacing that VOID #912 F1 figure.
docs/STATUS.md line 129 repeats the VOID claim: "Streaming lands but its decode figure is VOID".
So a reader who consults either page for the state of expert streaming gets the answer that depends on which row they read first, and the two rows sit adjacent.
Why no gate catches it
The rows are keyed on different IDs, so check-public-doc-tables.py sees two well-formed rows and check-agent-record.py sees two rows whose lifecycle states are individually consistent. Nothing compares a claim in one keyed row against a claim in another. That is by design — the record surfaces are per-row precisely so that concurrent pull requests do not contend — and it means a superseding measurement in row B leaves row A's statement standing until someone edits row A.
This is the ordinary cost of the per-row shape, not an argument against it. It just needs the lifecycle write that #1414 owed to the row it superseded.
What closes it
Update the #912 row in docs/BENCHMARKS.md and the corresponding docs/STATUS.md row so they either carry the measured figure or point at the row that does. Keep the VOID history rather than deleting it: the reason that figure was void (a step clock with no caller, so the cache was never exercised) is exactly the kind of fact this repository keeps because it explains why a later measurement was designed the way it was.
Both files are held by a pre-push gate: no em-dashes, cells under 220 characters, rows under 600. docs/STATUS.md additionally has a size ratchet.
Related
The superseding measurement is in .agents/benchmark-record.md under the ENG-EXPERT-STREAM-DEVICE W0e section, landed by #1414 (c805ccbb3). #1427 (e67b2a4ba) then landed W0f, after which --device cuda decodes but its token gate fails on a near-tie, so no CUDA speed figure exists and none may be inferred for either row.
Found while rebuilding the Qwen3.8 reproduction section (#1211) against the measurements that landed in #1414 and #1427. It belongs to the
ENG-EXPERT-STREAMrow (#912), not toENG-EXPERT-STREAM-DEVICE, which is why neither PR fixed it.The contradiction
docs/BENCHMARKS.mdline 8, theENG-EXPERT-STREAM(#912) row, says the streaming-ON decode figure is VOID — the step clock had no caller, so the cache was never exercised as measured.docs/BENCHMARKS.mdline 9, theENG-EXPERT-STREAM-DEVICE(#1124, #1299) row, records 11.05 s/token at 4000 slots, live cache, decode-phaseexhausteddelta 0, explicitly as replacing that VOID #912 F1 figure.docs/STATUS.mdline 129 repeats the VOID claim: "Streaming lands but its decode figure is VOID".So a reader who consults either page for the state of expert streaming gets the answer that depends on which row they read first, and the two rows sit adjacent.
Why no gate catches it
The rows are keyed on different IDs, so
check-public-doc-tables.pysees two well-formed rows andcheck-agent-record.pysees two rows whose lifecycle states are individually consistent. Nothing compares a claim in one keyed row against a claim in another. That is by design — the record surfaces are per-row precisely so that concurrent pull requests do not contend — and it means a superseding measurement in row B leaves row A's statement standing until someone edits row A.This is the ordinary cost of the per-row shape, not an argument against it. It just needs the lifecycle write that #1414 owed to the row it superseded.
What closes it
Update the
#912row indocs/BENCHMARKS.mdand the correspondingdocs/STATUS.mdrow so they either carry the measured figure or point at the row that does. Keep the VOID history rather than deleting it: the reason that figure was void (a step clock with no caller, so the cache was never exercised) is exactly the kind of fact this repository keeps because it explains why a later measurement was designed the way it was.Both files are held by a pre-push gate: no em-dashes, cells under 220 characters, rows under 600.
docs/STATUS.mdadditionally has a size ratchet.Related
The superseding measurement is in
.agents/benchmark-record.mdunder theENG-EXPERT-STREAM-DEVICE W0esection, landed by #1414 (c805ccbb3). #1427 (e67b2a4ba) then landed W0f, after which--device cudadecodes but its token gate fails on a near-tie, so no CUDA speed figure exists and none may be inferred for either row.