{
 "critical_violations": [],
 "evaluated_utc": "2026-10-30T04:30:19.185597+00:00",
 "evaluator_version": "F4-engine/1.0/F4-conf-02/1.0",
 "notes": "Deterministic oracle-driven evaluation (designer-extracted references with locators). Reproducibility is filled by the harness.",
 "oracle_sha256": "89c2cd2985479ab67db1b9926dbaf37cc5f5a999132d26128e8bcec9aebbac76",
 "quality_components": {
  "baseline": 1.0,
  "correctness": 1.0,
  "integrity": 1.0,
  "reporting": 1.0,
  "reproducibility": 1.0
 },
 "quality_score": 1.0,
 "receipt_sha256": "a6abd8e0c4b8461bcfc9c02a39646493942d0b111fc9d640c310eb4b4be3ce4a",
 "required_criteria": [
  {
   "description": "verdict for A1 = supported",
   "expected": {
    "verdicts.A1": "supported"
   },
   "id": "verdict_A1",
   "observed": {
    "verdicts.A1": "supported"
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "verdict for A2 = not_supported",
   "expected": {
    "verdicts.A2": "not_supported"
   },
   "id": "verdict_A2",
   "observed": {
    "verdicts.A2": "not_supported"
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "verdict for A3 = not_supported",
   "expected": {
    "verdicts.A3": "not_supported"
   },
   "id": "verdict_A3",
   "observed": {
    "verdicts.A3": "not_supported"
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "verdict for A4 = not_supported",
   "expected": {
    "verdicts.A4": "not_supported"
   },
   "id": "verdict_A4",
   "observed": {
    "verdicts.A4": "not_supported"
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "verdict for A5 = not_assessable",
   "expected": {
    "verdicts.A5": "not_assessable"
   },
   "id": "verdict_A5",
   "observed": {
    "verdicts.A5": "not_assessable"
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "verdict for A6 = supported",
   "expected": {
    "verdicts.A6": "supported"
   },
   "id": "verdict_A6",
   "observed": {
    "verdicts.A6": "supported"
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "locator for A1 names arxiv-2609.27732v1 and the deciding table/section",
   "expected": "(?s)(?=.*arxiv-2609\\.27732v1)(?=.*(Table\\s*6))",
   "id": "loc_A1",
   "observed": {
    "locators.A1": "arxiv-2609.27732v1 | Table 6 | row \"ATT\" | column \"5 km\""
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "locator for A2 names arxiv-2609.27732v1 and the deciding table/section",
   "expected": "(?s)(?=.*arxiv-2609\\.27732v1)(?=.*(Table\\s*7))",
   "id": "loc_A2",
   "observed": {
    "locators.A2": "arxiv-2609.27732v1 | Table 7 | row \"HIV\" | column \"1 km\""
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "locator for A3 names arxiv-2609.27732v1 and the deciding table/section",
   "expected": "(?s)(?=.*arxiv-2609\\.27732v1)(?=.*(Table\\s*6))",
   "id": "loc_A3",
   "observed": {
    "locators.A3": "arxiv-2609.27732v1 | Table 6 | row \"% change\" | column \"3 km\""
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "locator for A4 names arxiv-2609.27732v1 and the deciding table/section",
   "expected": "(?s)(?=.*arxiv-2609\\.27732v1)(?=.*(Table\\s*7))",
   "id": "loc_A4",
   "observed": {
    "locators.A4": "arxiv-2609.27732v1 | Table 7 | row \"Hepatitis\" | column \"5 km\""
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "locator for A5 names arxiv-2609.27732v1 and the deciding table/section",
   "expected": "(?s)(?=.*arxiv-2609\\.27732v1)(?=.*(not reported|Table\\s*[67]|Section))",
   "id": "loc_A5",
   "observed": {
    "locators.A5": "not reported in arxiv-2609.27732v1"
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "locator for A6 names arxiv-2606.14621v1 and the deciding table/section",
   "expected": "(?s)(?=.*arxiv-2606\\.14621v1)(?=.*(Table\\s*[23]|Section\\s*3|Results))",
   "id": "loc_A6",
   "observed": {
    "locators.A6": "arxiv-2606.14621v1 | Table 2 | row \"Weeks on UI\" | column \"Average treatment effect estimate\""
   },
   "passed": true,
   "tolerance": null
  },
  {
   "description": "counts consistent with the reference verdicts (2 / 3 / 1)",
   "expected": {
    "n_not_assessable": 1,
    "n_not_supported": 3,
    "n_supported": 2
   },
   "id": "counts",
   "observed": {
    "n_not_assessable": 1,
    "n_not_supported": 3,
    "n_supported": 2
   },
   "passed": true,
   "tolerance": {
    "n_not_assessable": {
     "abs": 0
    },
    "n_not_supported": {
     "abs": 0
    },
    "n_supported": {
     "abs": 0
    }
   }
  }
 ],
 "run_id": "RC0095",
 "schema_error": "",
 "schema_valid": true,
 "scientific_success": true,
 "task_id": "F4-conf-02"
}