{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-10-26-marker-gpu-trace",
  "complete": true,
  "identity": {
    "source_commit": "626f6d8abe8f987a49e09ceab24651684e28ef4f",
    "source_clean_before_lifecycle": true,
    "candidate_image": "sha256:789be89fc1f16c28673fad1dbb539974f7d363c6c04ee1372a22fd12b21090d3",
    "runtime_source": "local-inference-lab/vllm@8244d1c44348ae8a5bbfc9f1b607021652283ed4",
    "upstream_patch": "0a1276e5e3d9000488e7590cc0957edd709f2152"
  },
  "workload": {
    "question": "For each accepted second reasoning-close marker, was its exact verification row budget-forced or sampled without a budget force?",
    "maximum_probe_invocations": 1,
    "maximum_requests": 2,
    "retries": 0,
    "reasoning_budget_tokens": 64,
    "maximum_completion_tokens": 512,
    "requested_concurrency": 1,
    "temperature": 0.6,
    "top_p": 0.95
  },
  "measurement": {
    "path": "synthetic actual-CUDA preweight gates followed conditionally by one route-derived direct online C1 probe",
    "warm_cold_state": "not a performance population; synchronized tracing and startup state retained only as diagnostic context",
    "sample_count": 1,
    "statistics": []
  },
  "gates": [
    {
      "gate": "actual_cuda_preweight_marker_selftest",
      "status": "passed",
      "evidence": "startup-gates.json"
    },
    {
      "gate": "sampling_upstream_count",
      "status": "passed",
      "evidence": "startup-gates.json"
    },
    {
      "gate": "complete_worker_trace_native_response_reconciliation",
      "status": "unknown",
      "evidence": "trace-result.json"
    },
    {
      "gate": "exact_incumbent_operational_restoration",
      "status": "passed",
      "evidence": "restoration.json"
    },
    {
      "gate": "strict_json_functional_restoration",
      "status": "failed",
      "evidence": "baseline-smoke.json and restoration-smoke.json"
    }
  ],
  "headline_results": [
    {
      "metric": "native_candidate_requests",
      "value": 1,
      "unit": "requests",
      "boundary": "zero retries"
    },
    {
      "metric": "native_completion_tokens",
      "value": 75,
      "unit": "tokens",
      "boundary": "64 reasoning tokens; one close token at zero-based index 64"
    },
    {
      "metric": "answer_oracle",
      "value": true,
      "unit": "boolean",
      "boundary": "exact answer 9567,1085,10652"
    },
    {
      "metric": "worker_trace_tokens_per_rank",
      "value": 76,
      "unit": "tokens",
      "boundary": "first 75 match native response; trace-only suffix token 1654"
    }
  ],
  "failures": [
    {
      "gate": "complete_worker_trace_native_response_reconciliation",
      "status": "unknown",
      "reason": "GPU committed token sequence differs from native response"
    }
  ],
  "limitations": [
    "No second reasoning-close marker reproduced, so the target second-marker provenance question was not exercised.",
    "The matching budget store for the only observed close is supplemental because complete trace/native reconciliation failed.",
    "Synchronized tracing excludes every performance claim.",
    "One correct response is not broad model-quality or KV-cache evidence.",
    "Operational restoration passed, but functional acceptance remains incomplete because strict JSON failed before and after and the baseline raw response was not retained.",
    "Router custody is verified; router configuration byte parity is not claimed because mounted configuration bytes were not separately captured before the trial."
  ],
  "decision": {
    "evidence_labels": ["functional"],
    "decision_labels": [
      "no-promotion"
    ],
    "promotion_authorized": false,
    "next_gate": "offline post-stop reconciliation contract and a separately frozen repetition study"
  },
  "evidence": [
    "request.json",
    "response.json",
    "probe-result.json",
    "trace-result.json",
    "marker-trace-summary.json",
    "candidate-attestation.json",
    "runtime-candidate-before.json",
    "runtime-candidate-after.json",
    "startup-gates.json",
    "build.json",
    "serving-worker-traces.json",
    "independent-pre-live-review.json",
    "independent-live-review.json",
    "independent-restoration-review.json",
    "restoration.json",
    "closure.json",
    "runtime-before.json",
    "runtime-restored.json",
    "router-restored.json",
    "router-transition-restored.json",
    "mode-restored.json",
    "resources-final.json",
    "baseline-smoke.json",
    "restoration-smoke.json"
  ],
  "metrics_coverage": null,
  "comparison_evidence": null
}