Skip to content

GPU marker-trace evidence

This directory contains the sanitized evidence bundle for the dated finding. The native benchmark artifacts remain authoritative; the files below provide a consistent campaign-level index.

Campaign boundary

  • Campaign ID: 2026-10-26-marker-gpu-trace
  • Capability: focused text-serving GPU marker provenance diagnostic
  • Repository revision: 626f6d8abe8f987a49e09ceab24651684e28ef4f (clean executable source)
  • Evidence class: functional
  • Outcome: diagnostic inconclusive; operational restoration passed; functional acceptance incomplete
  • Decision labels: no-promotion
  • Promotion boundary: diagnostic evidence does not authorize promotion

The frozen question is whether each accepted second reasoning-close marker was budget-forced on its exact verification row or sampled without a budget force. The experiment preserves model, model revision, launch settings, MTP3, GPU allocation, TP2/DCP2, configured C4 capacity, request controls, and the prior split-top-p patch. Synchronized tracing excludes every performance claim. The final image directly composes the split-top-p and trace patches from the original immutable runtime image; the prior patched image is retained only as a reference comparison identity.

Common campaign artifacts

Workload and plan

One invocation permits at most two C1 requests, zero retries, a 120-second request timeout, and a 300-second request window. The first case uses a 64-token reasoning budget, 512-token completion cap, maximum reasoning effort, temperature 0.6, and top-p 0.95. The second case is allowed only if the first naturally closes below the boundary and answers correctly. See the frozen plan, source freeze, build, candidate attestation, and independent pre-live review.

Raw run evidence

The actual-CUDA preweight marker selftest and the frozen sampling gate passed before weights. The only request returned the correct answer and one closing marker at the 64-token boundary. Both TP ranks recorded a corresponding budget store, but the trace recorded 76 committed tokens while the native API returned the first 75, ending at its stop token. The frozen analyzer therefore returned unknown. No second marker reproduced, so the target second-marker question was not exercised. Exact operational restoration passed. Strict JSON failed before and after, so functional acceptance remains incomplete; the two smokes do not establish a quality regression.

Retained live evidence includes the request, response, native probe result, frozen analyzer result, complete public-safe worker records, replayable worker log, startup receipts, derived trace summary, independent post-probe review, and candidate runtime inspections before and after.

Decision and publication

The dated finding, final manifest, authoritative summary and projection ledger bind this closed packet. No compact platform copy was requested, so the publication-summary role is not applicable. The finding preserves the frozen unknown analyzer verdict, distinguishes the observed first-close store from the unexercised second-close question, and avoids counterfactual, performance, broader model/KV, or resolution claims. Router custody is restored, but router configuration byte parity is not claimed because mounted bytes were not captured before the trial.

Restoration evidence: restoration, closure, incumbent runtime before and restored, router custody, router transition, managed mode, resources, strict JSON smokes before and after, and the independent restoration review.

From the repository root, replay the frozen analyzer with:

python3 configs/runtime-patches/vllm/8244d1c4-marker-trace/analyze.py --logs docs/findings/2026-10-26-marker-gpu-trace-evidence/serving-worker-traces.log --startup-logs docs/findings/2026-10-26-marker-gpu-trace-evidence/startup-receipts.log --response docs/findings/2026-10-26-marker-gpu-trace-evidence/response.json --output /tmp/marker-trace-replay.json

The expected exit code is 1 and the expected status is unknown because the trace contains one token after the API-visible stop boundary. The replay hashes the sanitized response and logs, so its artifact hashes differ from the frozen native analyzer receipt while its gate outcome is unchanged. The first public replay receipt failed earlier at missing prompt_token_ids; it is retained rather than overwritten. The corrected replay receipt reaches the frozen sequence-mismatch error.