{
  "evaluation_kind": "historical_regression",
  "dataset_sha256": "9a3b873882bbf29b905d79c9448c3c287fda30323f93a6533b88dbd59b19f558",
  "graph_dev": 722,
  "graph_regression": 320,
  "llm_dev_ids": [
    "wide-001",
    "wide-002",
    "wide-041",
    "wide-042",
    "wide-080",
    "wide-081",
    "wide-082",
    "wide-104",
    "wide-113",
    "wide-114",
    "wide-119",
    "wide-137",
    "wide-138",
    "wide-139",
    "wide-141",
    "wide-142",
    "wide-143",
    "wide-145",
    "wide-155",
    "wide-159",
    "wide-161",
    "wide-163",
    "wide-169",
    "wide-170",
    "wide-178",
    "wide-181",
    "wide-193",
    "wide-194",
    "wide-209",
    "wide-217",
    "wide-218",
    "wide-232",
    "wide-236",
    "wide-239",
    "wide-249",
    "wide-250",
    "wide-283",
    "wide-289",
    "wide-290"
  ],
  "llm_dev_count": 39,
  "llm_final_regression": 320,
  "llm_final_dev_overlap": 39,
  "target_strict_accuracy": 0.92,
  "format_repair_limit": 1,
  "policy": "Do not tune after final freeze. Retain all failures. Same question and context for one format-only retry; never send gold to LLM.",
  "academic_preparation_policy": "Preparatory/planning headings do not establish an academic work or year.",
  "limitations": [
    "Known corpus, persons, templates and questions. Repeated measurement is regression, not independent TEST.",
    "Native OpenWebUI tool routing is point 4."
  ],
  "previous_iteration": {
    "graph": "v5: 319/320; consultation sentence misclassified as title",
    "llm_dev": "38/39; no JSON errors"
  }
}