{
  "evaluation_scope": "Separate local offline studies, not official challenge scores.",
  "studies": [
    {
      "id": "segment-refinement",
      "track": "Segment",
      "label": "Answer-guided temporal refinement",
      "baseline": 50.4,
      "improved": 55.0,
      "n": 1577,
      "subset": "Single-timestamp questions",
      "change": "One additional 64-frame pass within ±15 seconds of the first answer.",
      "boundary": "Earlier Segment checkpoint; this is not an overall track score.",
      "source": "Segment inference temporal-cascade comments; archived spotlight fact check."
    },
    {
      "id": "segment-question-window",
      "track": "Segment",
      "label": "Question-anchored evidence windows",
      "baseline": 72.7,
      "improved": 79.1,
      "n": 422,
      "subset": "Questions that state a timestamp",
      "change": "Spend the same 64 frames inside ±30-second question-anchored windows.",
      "boundary": "A separate question population from the refinement study.",
      "source": "Segment resources/question_window.py, recorded window-width study."
    },
    {
      "id": "procedure-search",
      "track": "Procedure",
      "label": "Coarse-to-fine temporal search",
      "baseline": 6.8,
      "improved": 35.1,
      "n": 746,
      "subset": "Single-timestamp questions",
      "change": "Three passes: full procedure → 1,200 seconds → 100 seconds, at 128/128/64 frames.",
      "boundary": "Single-checkpoint search study before the later W64/NW2 routing split.",
      "source": "Procedure inference temporal-cascade comments; archived spotlight fact check."
    }
  ],
  "pre_evaluation": {
    "segment": 0.5656564984886887,
    "procedure_technical": 0.4112059081777602,
    "procedure_clinical": 0.5572717019874663,
    "source": "The two repositories' provenance/release.json records."
  }
}