{
  "experiment": "exp-004-pooled-block-gate",
  "date": "2026-10-03",
  "fill": {
    "client_calls": 2073,
    "usd": 1.238223,
    "ceiling_usd": 2.0,
    "still_missing": 0,
    "dry_run_estimate_usd": 5.582,
    "note": "the dry-run estimator over-priced 4.5x (no prompt-cache pricing); measured spend is the meter's"
  },
  "run": {
    "units": 501,
    "client_calls": 0,
    "usd": 0.0,
    "replayed": 2509,
    "replay_miss": 0
  },
  "verdict": "inconclusive",
  "reason": "no t_block on any outer fold, pooled or per family; B offers no BLOCK",
  "instrument_check": {
    "per_family": {
      "access_control": {
        "n": 66,
        "distinct_scores": 41,
        "unsure": 5,
        "pairs": 28,
        "within_pair_wins": 22,
        "ties": 1,
        "positives_above_first_negative": 12
      },
      "deserialization": {
        "n": 64,
        "distinct_scores": 38,
        "unsure": 7,
        "pairs": 25,
        "within_pair_wins": 18,
        "ties": 4,
        "positives_above_first_negative": 5
      },
      "injection": {
        "n": 149,
        "distinct_scores": 52,
        "unsure": 54,
        "pairs": 53,
        "within_pair_wins": 30,
        "ties": 17,
        "positives_above_first_negative": 6
      },
      "output_encoding": {
        "n": 69,
        "distinct_scores": 31,
        "unsure": 32,
        "pairs": 28,
        "within_pair_wins": 18,
        "ties": 6,
        "positives_above_first_negative": 2
      },
      "path": {
        "n": 66,
        "distinct_scores": 43,
        "unsure": 15,
        "pairs": 29,
        "within_pair_wins": 25,
        "ties": 3,
        "positives_above_first_negative": 13
      },
      "untrusted_destination": {
        "n": 87,
        "distinct_scores": 36,
        "unsure": 33,
        "pairs": 32,
        "within_pair_wins": 18,
        "ties": 4,
        "positives_above_first_negative": 3
      }
    },
    "pooled_raw_top30_labels": "111111111110111110111111110111",
    "reading": "scores are real (31-52 distinct values per family); pooled raw top 30 holds 27 positives (0.90), first negative at rank 12; the Wilson bound needs >= 73 flags with no error, so precision at the top, not data volume alone, now binds"
  }
}