{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "glm-sampling-upstream-tolerance-20261025",
  "complete": true,
  "identity": {
    "source_head": "80dacf71b72d9760cf539d7cbe8e9c8a88c25647",
    "model": {
      "repository": "nvidia/GLM-5.3-Flash-FP4-Local",
      "revision": "76348ddb2be2e35e39ecfc00e75dc07703f4f3d5",
      "public_served_label": "served-model-1",
      "weight_format": "FP4",
      "kv_dtype": "FP8",
      "context": 327680,
      "max_concurrency": 4,
      "tensor_parallel": 2,
      "decode_context_parallel": 2,
      "speculative_tokens": 3
    },
    "runtime": {
      "base_revision": "8244d1c44348ae8a5bbfc9f1b607021652283ed4",
      "upstream_patch_commit": "0a1276e5e3d9000488e7590cc0957edd709f2152",
      "base_image": "sha256:745feb916ee816f5409fd58386114be4f22fbede7f9c0e0a7b8c4fe8dac0fb29",
      "candidate_image": "sha256:41c4926b08c7298fb63bc7f1204926c4f95ce9083ff610069ab442141213b1aa",
      "baseline_sampler_sha256": "0b5fc337058e22bc28a8530ccec904d6ccd89b9548b02955c657ac429b1669d2",
      "patched_sampler_sha256": "2b5305c406d8495e515f048ce0322c7ae5946a49e5f27bd5bbbf30a1e1fea0e6"
    }
  },
  "workload": {
    "scope": "Focused functional kernel and single model reasoning-boundary diagnostic; no performance comparison",
    "policy": "upstream-count-v1",
    "plan": "plan.json",
    "probe_plan": "probe-plan.json"
  },
  "measurement": {
    "path": "synthetic CUDA kernels, followed by one isolated C1 direct-endpoint request through a route-derived in-memory relay",
    "warm_cold_state": "Fresh candidate load; JIT/cache uncontrolled; synchronized diagnostic traces",
    "forced_cases": 8,
    "ordinary_cases": 8,
    "ordinary_rows": 272,
    "candidate_requests": 1
  },
  "kernel": {
    "gate": "passed",
    "failed_rows": 0,
    "failed_native_batches": 0,
    "stock_patch_equal_rows": 271,
    "baseline_exact_binary64_reference_rows": 36,
    "patched_exact_binary64_reference_rows": 35,
    "peak_tensor_bytes": 519289856,
    "prior_one_token_difference_preserved": true
  },
  "boundary": {
    "schema": "anvil-serving.sampling-boundary-summary/v1",
    "derived_from": [
      "response.json",
      "probe-result.json",
      "serving-worker-traces.json"
    ],
    "requests": 1,
    "finish_reason": "stop",
    "completion_tokens": 137,
    "native_reasoning_tokens": 64,
    "closing_token": 154842,
    "closing_token_positions_zero_based": [
      64,
      126
    ],
    "token_zero_count": 0,
    "exact_answer_oracle": false,
    "boundary": "invalid_repeated_or_missing_close",
    "numeric_solution_present": true,
    "numeric_solution_presence_is_not_strict_oracle_pass": true,
    "candidate_worker_trace_receipts": 2,
    "kernel_path_evidence": "observed synchronized split-mask dispatch",
    "full_reasoning_limit_gate": "failed",
    "promotion_authorized": false
  },
  "gates": [
    {
      "name": "synthetic-kernel",
      "status": "passed"
    },
    {
      "name": "candidate-source-identity",
      "status": "passed"
    },
    {
      "name": "serving-worker-dispatch",
      "status": "observed"
    },
    {
      "name": "reasoning-boundary",
      "status": "failed"
    },
    {
      "name": "strict-answer",
      "status": "failed"
    },
    {
      "name": "baseline-json-smoke",
      "status": "failed"
    },
    {
      "name": "operational-restoration",
      "status": "passed"
    },
    {
      "name": "functional-restoration-smoke",
      "status": "failed"
    }
  ],
  "headline_results": [
    {
      "name": "forced_cases_passed",
      "value": 8
    },
    {
      "name": "ordinary_rows_checked",
      "value": 272
    },
    {
      "name": "stock_patch_equal_rows",
      "value": 271
    },
    {
      "name": "candidate_requests",
      "value": 1
    },
    {
      "name": "closing_token_positions_zero_based",
      "value": [
        64,
        126
      ]
    },
    {
      "name": "token_zero_count",
      "value": 0
    }
  ],
  "failures": [
    "Repeated reasoning-end marker and duplicated answer/prose fail strict model gates.",
    "Pretrial strict JSON smoke contains prose.",
    "Matched restored-incumbent strict JSON smoke also fails; no quality regression established."
  ],
  "limitations": [
    "Single model request; no retry or causal isolation of repeated-marker behavior.",
    "Dispatch receipts are not per-token provenance or profiler-level evidence.",
    "Trace vocabulary 154880 versus tokenizer/fixture 154856; padding explanation not independently verified.",
    "Native probe keeps runtime_verified=false and drain unknown; later inspector and attestation are separate.",
    "No performance, capacity, broad quality or promotion result."
  ],
  "restoration": "operational_verified_existing_json_failure_reproduced",
  "decision": {
    "evidence_labels": [
      "functional"
    ],
    "decision_labels": [
      "no-promotion"
    ],
    "promotion_authorized": false,
    "next_gate": "Offline investigation of repeated-close MTP/thinking-budget/parser state; preserve failed strict outcome and matched JSON-smoke quality limit."
  },
  "promoted": false,
  "ignis": "hold",
  "evidence": [
    "kernel-receipts.log",
    "kernel-summary.json",
    "request.json",
    "response.json",
    "probe-result.json",
    "boundary-summary.json",
    "serving-worker-traces.json",
    "candidate-attestation.json",
    "runtime-candidate-before.json",
    "runtime-candidate-after.json",
    "baseline-smoke.json",
    "restoration.json",
    "closure.json",
    "restoration-smoke.json",
    "independent-kernel-review.json",
    "independent-live-review.json",
    "trace-freshness.json"
  ],
  "quality_regression_established": false
}