{
  "schema": "anvil-serving.controlled-ab-diagnostic-plan/v1",
  "created_at": "2026-10-17T08:23:37.586923+00:00",
  "authorization": "User explicitly requested retry/debug of candidate to assess configuration-related failure after first restoration.",
  "original_result": "candidate r2 context17/18 remains failed and frozen",
  "design": {
    "r2": [
      "original full18 context replay",
      "fixed selected failure repeat1",
      "fixed selected failure repeat2"
    ],
    "all_jobs_repetitions": 1,
    "seed_note": "Builder seed17, repetition0; repetitions3 would change fixtures. Generation sampling remains temperature1/top_p.95 with no request seed.",
    "r3_condition": "At least one of the3new r2 target observations fails",
    "r3_runtime_delta": "remove only --speculative-config; separate recipe/container/served names and descriptive metadata; all other flags/env/image/model/cgroup/GPU order unchanged",
    "r3_jobs": "same full18 plus2 targeted jobs",
    "exact_fixture_required": [
      "case_id",
      "answer_sha256",
      "calibration ratio4.832231202547146",
      "reconstructed prompt SHA a8d952496fe17525e8c404823f107e0c2f121775cba5ad7ae3624e8fc05d59bd",
      "observed_prompt_tokens316249"
    ],
    "wire_hash_limit": "Original runner does not retain submitted prompt/hash; matching reconstruction is not cryptographic wire-byte proof.",
    "history": "Full suite and subsequent isolated warm repeats reported separately; no cache flush."
  },
  "bounds": {
    "total_seconds": 10800,
    "restoration_reserve_seconds": 4200,
    "request_cutoff_seconds": 6600,
    "per_job_seconds": 900,
    "lifecycle_seconds": 3600
  },
  "stop": "Complete all3 precommitted jobs regardless of quality pass/fail; abort on timeout, incomplete count, transport/runtime/OOM/fatal error, identity drift or reserve boundary.",
  "review": {
    "reviewer": "independent-adversarial-reviewer",
    "decision": "accept",
    "conditions_applied": true
  },
  "excluded": [
    "timing",
    "SWE",
    "promotion",
    "unrelated tuning",
    "host swap changes"
  ],
  "restoration": "Exact retained original r11 through previously verified managed restore-retained; direct and authenticated routed JSON after readmission",
  "promoted": false
}