{
 "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"
 ]
}
