{
 "schema": "cain.e127b.compute_matched.v1",
 "at": "2026-11-21T04:48:40Z",
 "entity": "FIRME AI CORPORATION",
 "experiment": "E127-B compute-matched control (cm_full gets the same evaluations as cv_filter_full)",
 "preregistration": {
  "schema": "cain.e127b.preregistration.v1",
  "frozen_at": "2026-11-21T04:47:39Z",
  "title": "E127-B \u2014 is the E127-A gain due to cross-validation or to more compute?",
  "motivation": "E127-A's winning arm evaluated the whole hypothesis space while the controls used 400 evaluations. This experiment gives the random-order control the SAME number of hypothesis evaluations per family.",
  "intervention": "cm_full: random-order search with budget = len(hypotheses for that family). Same hypothesis space, same evaluation count as cv_filter_full.",
  "splits": {
   "development_variants": [
    7,
    8,
    9
   ],
   "sealed_test_variants": [
    15,
    16,
    17,
    18,
    19
   ]
  },
  "arms": [
   "baseline_current",
   "compute_matched_400",
   "cv_filter_full",
   "cm_full"
  ],
  "primary_endpoint": "sealed accuracy cv_filter_full vs cm_full on variants 15-19",
  "success_threshold": "cv_filter_full beats cm_full by >= 5.0 pts with exact McNemar p < 0.05",
  "failure_rule": "if below threshold, the E127-A gain is attributable to compute and will be reported as such",
  "budget": "local CPU only, $0",
  "limits": [
   "constructed generator",
   "paired per-instance outcomes",
   "no external benchmark",
   "no human review"
  ]
 },
 "development": {
  "baseline_current": 0.8083333333333333,
  "compute_matched_400": 0.8,
  "cv_filter_full": 0.9916666666666667,
  "cm_full": 1.0,
  "n_instances": 120,
  "cv_vs_cm_full": {
   "cv_only": 0,
   "cm_only": 1,
   "p": 1.0
  }
 },
 "sealed_test": {
  "baseline_current": 0.8,
  "compute_matched_400": 0.805,
  "cv_filter_full": 0.98,
  "cm_full": 0.975,
  "n_instances": 200,
  "cv_vs_cm_full": {
   "cv_only": 1,
   "cm_only": 0,
   "p": 1.0
  }
 },
 "delta_cv_vs_cm_full": 0.0050000000000000044,
 "verdict": "GAIN_WAS_COMPUTE",
 "limits": [
  "constructed generator",
  "paired per-instance outcomes",
  "no external benchmark",
  "no human review"
 ],
 "agi_status": "NOT ACHIEVED"
}
