Independent APC promotion review — 2026-09-19

Recommendation: promote r10 APC within the tested configuration, subject to the separate human promotion gate. human_gate_required=true; promoted=false.

Independence: campaign lead/evidence author is reported as GPT-6 Astra; the evaluated endpoint is GLM-5.3-Flash r9 versus r10 APC; reviewer is GPT-5.6 Sol. Neither author nor evaluated model is Sol.

Evidence verdict

The frozen 32-request C4 cells are sufficient to stop capacity testing. Both arms completed 32/32 with strict output adherence and request canaries. On the shared 32K cell, r10 reduced mean visible TTFT 67.84%, reduced time-to-first-output 73.15%, and increased requests/second 216.55%. Isolated engine counters recorded a 75.58% prefix-cache hit fraction, providing causal APC evidence. On the unique 32K cell, requests/second regressed 2.39%, inside the frozen 5% limit; mean TTFT rose 4.02%, time-to-first-output 6.22%, and E2E 2.73% while completion volume rose 5.42%. This is an acceptable bounded cost under the stated throughput gate, not evidence of a universal latency improvement. Retain the negative eight-request scout as variance/cold-path evidence.

Candidate correctness evidence is broad enough for this runtime-only claim: direct preflight passed; candidate and restored-control quality checks each passed 12/12 across three repetitions with typed control evidence from the actual pinned template rendering native Max effort; candidate image/OCR and both arms' eight-image tests passed; 309,422-token needle and 110,760-token tool-call checks passed; C4 and the 327,680 configured limit were retained; and status evidence shows no OOM or swap event under the asserted 1 MiB candidate bound. This does not establish broad intelligence parity or exhaustive full-context C4 behavior.

Gate and evidence status

Quality, functional, performance, containment, and restoration gates are closed. restoration.json verifies the exact r9 recipe/image/checkpoint, direct and authenticated routed preflight 6/6, route readmission, unchanged router configuration, and absence of the candidate container/listener. The corrected systemd receipt shows the wrapper remained active (exited) from September 13 and was not activated by this campaign. The candidate unique finalist lacks an embedded configuration fingerprint, but recipe-bindings.json binds its retained artifact hash and running-identity receipt to the exact normalized r10 recipe hash; this is adequate for this campaign while remaining a native-artifact limitation.

Recommend human-approved bounded promotion of r10 APC with the existing model, checkpoint, image, precision, 327K configuration, C4, vision-eight, and no-speculation settings unchanged. Keep promoted=false until that human-gated operation succeeds. Roll back on unique-throughput regression above 5%, correctness failure, OOM/containment failure, or loss of direct or routed acceptance. No further pre-promotion performance, vision, context, or quality tests are needed.

Publication review

The final public finding and companion evidence bundle are accepted for publication finalization. Their performance, quality, limitation, restoration, and human-gate claims match the private evidence. Direct and routed endpoints are synthetic; the reviewed public files contain no operator path, host identity, GPU UUID, container ID, or private network identity. Finalization must copy this corrected review and regenerate the artifact hashes afterward.