{
 "pooled": {
  "C": {
   "chunk_mae": 10.5024,
   "first_mae": 8.5606,
   "pred_key": "pred:bijou@40000"
  },
  "A-s0": {
   "chunk_mae": 7.7966,
   "first_mae": 3.9422,
   "pred_key": "pred:bijou@40000"
  }
 },
 "read1_primary_C_minus_As0": {
  "definition": "paired per-frame chunk_mae, C - A-s0, core frames",
  "mean": 2.64001,
  "median": 1.27433,
  "ci95": [
   2.54793,
   2.73598
  ],
  "c_win_rate": 0.2392,
  "n_frames": 17204,
  "pooled_dchunk": 2.7058,
  "most_influential_repos": [
   {
    "repo": "dopaul/1500_chess_moves",
    "n_excl": 600,
    "mean_without": 2.57774
   },
   {
    "repo": "Dongkkka/vla_total_dataset_test4",
    "n_excl": 72,
    "mean_without": 2.58233
   },
   {
    "repo": "jlesein/TestBoulon7",
    "n_excl": 160,
    "mean_without": 2.609
   },
   {
    "repo": "Hale-Sage/merge_data",
    "n_excl": 260,
    "mean_without": 2.66556
   },
   {
    "repo": "willnorris/bbox-2",
    "n_excl": 16,
    "mean_without": 2.61632
   }
  ]
 },
 "read2_reliance": {
  "subset_rows": 4301,
  "masked_first": 8.1066,
  "masked_chunk": 11.772,
  "intact_first_on_subset": 8.1243,
  "intact_chunk_on_subset": 11.7776,
  "delta_first": {
   "mean": -0.01771,
   "ci95": [
    -0.12237,
    0.0831
   ]
  },
  "delta_chunk": {
   "mean": -0.00556,
   "ci95": [
    -0.066,
    0.05362
   ]
  },
  "sanity_gate": {
   "rule": "delta_first < 5.0 (vs A-s0's +19.95)",
   "passed": true
  },
  "capability": {
   "masked_first": 8.1066,
   "class": "partial",
   "comparators_same_rows": {
    "intact_state_copy_first": 2.4316,
    "As0_masked_first": 23.8154,
    "B_masked_first": 24.0783
   },
   "thresholds": {
    "qualitative_first_lt": 2.4316,
    "strong_lt": 6.0,
    "failed_ge": 15.0
   }
  }
 },
 "read3_grounding": {
  "c_first_mae": 8.5606,
  "as0_first_mae": 3.9422,
  "delta_first_vs_as0": 4.6184,
  "state_copy_floor": 2.6202,
  "joint_interpretation": "no first_mae gain to attribute"
 },
 "E4_prior_band": {
  "band": [
   7.65,
   8.3
  ],
  "c_chunk_mae": 10.5024,
  "inside": false
 },
 "decision": {
  "read1_mean": 2.64001,
  "band": 0.15,
  "sanity_gate_passed": true,
  "capability_class": "partial",
  "probe_final": null,
  "probe_gate_ok": true,
  "primary_verdict": "COSTS: p=0.8 costs actions beyond the band",
  "adoption": "ADOPT NOTHING",
  "branch": "p=0.3 screen is the one sanctioned follow-up (mechanism worked: reliance collapsed)"
 }
}