{
  "date": "2026-10-03",
  "scope": "Frozen v1.1 development snapshot; not a rerun of the current production serving revision. Opened static/GQA subsets; fresh video comparison previously frozen.",
  "rows": [
    {
      "scene": "Static yes/no",
      "n": 154,
      "spark_correct": 144,
      "clef_flash_correct": 145
    },
    {
      "scene": "Color",
      "n": 38,
      "spark_correct": 38,
      "clef_flash_correct": 38
    },
    {
      "scene": "Counting",
      "n": 48,
      "spark_correct": 44,
      "clef_flash_correct": 43
    },
    {
      "scene": "Scene semantics",
      "n": 80,
      "spark_correct": 69,
      "clef_flash_correct": 66
    },
    {
      "scene": "Static set \u00b7 total",
      "n": 320,
      "spark_correct": 295,
      "clef_flash_correct": 292
    },
    {
      "scene": "GQA object existence",
      "n": 71,
      "spark_correct": 55,
      "clef_flash_correct": 57
    },
    {
      "scene": "GQA attributes",
      "n": 228,
      "spark_correct": 176,
      "clef_flash_correct": 172
    },
    {
      "scene": "GQA spatial relations",
      "n": 95,
      "spark_correct": 81,
      "clef_flash_correct": 73
    },
    {
      "scene": "GQA global (small diagnostic)",
      "n": 6,
      "spark_correct": 5,
      "clef_flash_correct": 5
    },
    {
      "scene": "GQA subset \u00b7 total",
      "n": 400,
      "spark_correct": 317,
      "clef_flash_correct": 307
    },
    {
      "scene": "Fresh video questions",
      "n": 120,
      "spark_correct": 65,
      "clef_flash_correct": 68
    }
  ],
  "method": "Same question/image pairs for 720 static/GQA items. Provider preprocessing may differ. Existing Spark results paired with later Clef calls, not simultaneous runs. One Spark GQA timeout and one Clef static authorization error remain in denominators. Aggregate rows overlap the subgroup rows.",
  "video_note": "120 questions clustered by source video; paired cluster 95% interval for Spark minus Clef: -10 to +5 percentage points. No equivalence claim.",
  "latency": {
    "scope": "Eight opened gesture clips, Colab T4, median observed times; no labeled gesture-accuracy claim. Prepare and run are two HTTPS calls.",
    "image_origin_ms": 669,
    "four_frame_origin_ms": 977,
    "image_client_https_ms": 1255,
    "four_frame_client_https_ms": 1632
  }
}