{
 "adversarial_audits": {
  "rejected": 7,
  "run": 7
 },
 "by_state": {
  "PARTIAL": 7,
  "PREREGISTERED": 23,
  "REJECTED": 7,
  "VERIFIED": 513
 },
 "excluded_runs": 0,
 "exported_at": "2026-11-22T02:00:43Z",
 "families": {
  "calibration_shift": {
   "ci95_across_experiments": [
    -0.016541,
    0.002209
   ],
   "domain": "metacognition / calibration",
   "experiments": 26,
   "holm_significant_global": 2,
   "hypothesis": "Target-domain ECE of the shift-aware method is lower than plain temperature scaling.",
   "lab": "AGI_LAB",
   "ladder_level": "L3",
   "mean_effect_across_experiments": -0.007166,
   "note": "experiments in a family share the hypothesis but use independent seeds and grid cells; the across-experiment interval treats each experiment as one unit",
   "outcomes": {
    "CONTRADICTED": 17,
    "INCONCLUSIVE": 2,
    "SUPPORTED": 7
   },
   "primary_tests": 21,
   "replication_agreement": 3,
   "replications": 5
  },
  "causal_backdoor": {
   "ci95_across_experiments": [
    0.246916,
    0.45035
   ],
   "domain": "causal inference",
   "experiments": 33,
   "holm_significant_global": 16,
   "hypothesis": "Mean |ATE error| of the treatment is lower than the naive regression on paired SCM instances.",
   "lab": "AGI_LAB",
   "ladder_level": "L2",
   "mean_effect_across_experiments": 0.348633,
   "note": "experiments in a family share the hypothesis but use independent seeds and grid cells; the across-experiment interval treats each experiment as one unit",
   "outcomes": {
    "INCONCLUSIVE": 11,
    "SUPPORTED": 22
   },
   "primary_tests": 31,
   "replication_agreement": 1,
   "replications": 2
  },
  "continual_drift": {
   "ci95_across_experiments": [
    0.026114,
    0.0918
   ],
   "domain": "continual learning / memory consolidation",
   "experiments": 15,
   "holm_significant_global": 2,
   "hypothesis": "Query accuracy of the change-point memory is higher than append-only majority memory.",
   "lab": "AGI_LAB",
   "ladder_level": "L3",
   "mean_effect_across_experiments": 0.058957,
   "note": "experiments in a family share the hypothesis but use independent seeds and grid cells; the across-experiment interval treats each experiment as one unit",
   "outcomes": {
    "CONTRADICTED": 3,
    "INCONCLUSIVE": 2,
    "SUPPORTED": 10
   },
   "primary_tests": 10,
   "replication_agreement": 5,
   "replications": 5
  },
  "goal_discovery": {
   "ci95_across_experiments": [
    0.02978,
    0.080532
   ],
   "domain": "goal discovery / active experimentation",
   "experiments": 16,
   "holm_significant_global": 1,
   "hypothesis": "Held-out reward prediction accuracy is higher with hypothesis-splitting probes than random probes at equal budget.",
   "lab": "AGI_LAB",
   "ladder_level": "L3",
   "mean_effect_across_experiments": 0.055156,
   "note": "experiments in a family share the hypothesis but use independent seeds and grid cells; the across-experiment interval treats each experiment as one unit",
   "outcomes": {
    "CONTRADICTED": 2,
    "INCONCLUSIVE": 6,
    "SUPPORTED": 8
   },
   "primary_tests": 15,
   "replication_agreement": 1,
   "replications": 1
  },
  "rsi_search": {
   "ci95_across_experiments": [
    -0.009896,
    0.001694
   ],
   "domain": "recursive algorithmic improvement",
   "experiments": 317,
   "holm_significant_global": 1,
   "hypothesis": "dev-selected child beats parent on held-out tasks",
   "lab": "RSI_LAB",
   "ladder_level": "L2",
   "mean_effect_across_experiments": -0.004101,
   "note": "experiments in a family share the hypothesis but use independent seeds and grid cells; the across-experiment interval treats each experiment as one unit",
   "outcomes": {
    "ACCEPTED": 11,
    "REJECTED_CANDIDATE": 306
   },
   "primary_tests": 317,
   "replication_agreement": 0,
   "replications": 0
  },
  "safety_killswitch": {
   "ci95_across_experiments": [
    35.245377,
    49.310179
   ],
   "domain": "safety and kill-switch verification",
   "experiments": 3,
   "holm_significant_global": 0,
   "hypothesis": "The gated arm produces fewer bad effects than the ungated control; the gated arm produces zero.",
   "lab": "GOVERNANCE_LAB",
   "ladder_level": "n/a",
   "mean_effect_across_experiments": 42.277778,
   "note": "experiments in a family share the hypothesis but use independent seeds and grid cells; the across-experiment interval treats each experiment as one unit",
   "outcomes": {
    "SUPPORTED": 3
   },
   "primary_tests": 2,
   "replication_agreement": 1,
   "replications": 1
  },
  "world_model_grid": {
   "ci95_across_experiments": [
    0.003217,
    0.007462
   ],
   "domain": "world modeling / interactive exploration",
   "experiments": 110,
   "holm_significant_global": 0,
   "hypothesis": "Next-state prediction accuracy over reachable (state, action) pairs is higher for curiosity than for random exploration.",
   "lab": "AGI_LAB",
   "ladder_level": "L4",
   "mean_effect_across_experiments": 0.005339,
   "note": "experiments in a family share the hypothesis but use independent seeds and grid cells; the across-experiment interval treats each experiment as one unit",
   "outcomes": {
    "CONTRADICTED": 32,
    "INCONCLUSIVE": 64,
    "SUPPORTED": 14
   },
   "primary_tests": 100,
   "replication_agreement": 5,
   "replications": 10
  }
 },
 "multiple_comparisons": {
  "method": "Holm\u2013Bonferroni across all primary experiment tests",
  "significant": 22,
  "tests": 496
 },
 "retries_counted_once": true
}
