{
  "calibration_outcomes_accessed": false,
  "changes": [
    "Source-only canonicalizer takes answer, question and task, never gold or model identity.",
    "Single-choice labelled text must agree exactly with its source option; text alone must resolve to one label.",
    "Truth triples enforce the source vocabulary and exact arity, with optional one balanced angle wrapper.",
    "Exact Decimal equality replaces float equality for numeric equivalence.",
    "Original raw official score and literal score remain diagnostics."
  ],
  "created_utc": "2026-10-22T09:26:45.524349+00:00",
  "final_outcomes_collected": false,
  "final_sensitivity": "Replay identical frozen policies additionally omitting source-ambiguous final IDs051/052; primary fixed population remains975. No subset selection on outcomes.",
  "format": "bbeh-grading-semantics-amendment-v2",
  "next_grades_prefix": "router-bbeh-recipes-20261021/grades-learn-v2-001",
  "reason": "A train/development raw-answer audit found unambiguous single-choice label plus exact source option text and source-demonstrated truth tuple delimiters rejected by the raw upstream grader.",
  "source_only_ambiguity_ids": [
    "bbeh:disambiguation_qa:018",
    "bbeh:disambiguation_qa:051",
    "bbeh:disambiguation_qa:052"
  ],
  "source_sha256": {
    "collection/grade.py": "b7e0da0e14b5e94c13b31f9fe70415752911bf03b69f394e6016f3d9a9717d24",
    "collection/grade_v2.py": "cd5d53a22b024406eecf18de92d24175a287e482dc63d58be6ec01b7a23cbd08",
    "collection/normalization_review/source-audit.json": "bf618b8be32a6391094b984919f802c8f0dd5c5c98417e8475764c56ee579cdd",
    "collection/normalize.py": "258a60874b11b5bbeefe2a4fe304dfcc2da341eba995258262ce1e5931ef5772",
    "collection/test_grade_v2.py": "41bdcc9004fe63db2317a3473a87d62aff64048939c4cde939992043800f002c",
    "collection/test_normalize.py": "fb74c8fe868d0e1a704c45da301db95b094314d3add6d311828ff171ef47a336",
    "operators/grade_answers_v2.py": "06f576aee5eedd9afe40c16666330cc9f69f1345a797b3d8a03922c4006155ba"
  },
  "superseded": [
    "grades-learn-001",
    "native/full-fit-001",
    "embedding-fit-001",
    "embedding-development-gross-001"
  ],
  "unchanged": [
    "3918 source population and family roles",
    "All original answer requests, settings, physical responses and charges",
    "All embedding vectors and encoder configuration",
    "Native and embedding hyperparameter candidate sets, selection loss and seed"
  ],
  "validation": "15 root regression/lineage tests and10 independent source-only adversarial tests passed; all4520 public source questions reviewed for formats."
}