{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-10-23-flare-state-diagnostics",
  "complete": true,
  "identity": {
    "model": "GLM-5.3-Flash",
    "source_revision": "47581ccac82660953a5b59cb04b2750d64edc97d",
    "candidate_patch": "b132065b27532028bf2a02e754e86553ea846281f616f22ce74e89442ff0ec76",
    "image_sha256": "d8fb208298fd3dd3f6b2b631148b01467a9449a9dcd1fef7dca304714bcf9411",
    "topology": "TP2; fixed 32 GiB cache pool"
  },
  "workload": {
    "protected_bracket": "small12 24 and pageheavy 12 measured requests per arm; A/B/B/A",
    "trace": "small12/pageheavy A/B captures; 72 measured requests, 144 scheduler samples"
  },
  "measurement": {
    "path": "online direct protected diagnostic; routed probes were restoration-only",
    "warm_cold_state": "matched warmup/prime phases retained; trace measured phase only",
    "sample_count": 264,
    "statistics": [
      "per-incarnation medians and nearest-rank p95 retained in analysis/protected-abba-2.json",
      "trace summary statistics are descriptive only",
      "No chart pack: the matched ABBA results remain performance-ineligible, so a chart would not support a qualified comparison."
    ]
  },
  "gates": [
    {
      "name": "protected matched population",
      "status": "passed diagnostic",
      "evidence": "analysis/protected-abba-2.json"
    },
    {
      "name": "native state roundtrip",
      "status": "passed reduced geometry only",
      "evidence": "analysis/state-logprob.json"
    },
    {
      "name": "trace completeness",
      "status": "passed diagnostic",
      "evidence": "analysis/trace-paired-1/trace-paired-1-summary.json"
    },
    {
      "name": "numerical qualification",
      "status": "failed/open",
      "reason": "no full-logit or independent teacher-forced oracle; prior exact-output gate remains unresolved"
    },
    {
      "name": "performance qualification",
      "status": "held",
      "reason": "protected regression threshold and causal attribution remain unresolved"
    }
  ],
  "headline_results": [
    {
      "name": "protected bracket completion",
      "value": "264/264",
      "condition": "24 warmups, 96 prime, 144 measured; failures=0"
    },
    {
      "name": "trace coverage",
      "value": "72/72 HTTP requests; 144 samples",
      "condition": "four complete 30-second TP2 captures; no drops or boundary exclusions"
    }
  ],
  "failures": [
    "protected ABBA-1 interrupted before timed phases",
    "state Window 1 native failure retained; two bounded repairs followed",
    "initial full test stopped after 51 passes because /tmp had a .git ancestor",
    "initial .venv Ruff invocation lacked Ruff"
  ],
  "limitations": [
    "A/B graph-shape mixtures differ",
    "no HTTP-to-slot/generation mapping",
    "trace cannot identify a production bottleneck or causal arm effect",
    "state oracle uses reduced geometry and is not logits or host-offload proof",
    "no performance qualification or promotion"
  ],
  "decision": {
    "evidence_labels": [
      "compatibility-only"
    ],
    "decision_labels": [
      "no-promotion"
    ],
    "promotion_authorized": false,
    "next_gate": "Bind HTTP requests to scheduler slots/generations and admission/host-index spans separately from timing; bind a raw-logit semantic package."
  },
  "evidence": [
    "report.md",
    "analysis/protected-abba-2.json",
    "analysis/state-logprob.json",
    "analysis/trace-paired-1/trace-paired-1-summary.json",
    "archives/reviews.tar.gz#review/final-trace-causal-review.md"
  ]
}