{
 "schema": "cain.e127.preregistration.v1",
 "frozen_at": "2026-11-21T04:41:18Z",
 "title": "E127-A \u2014 cross-validation hypothesis selection to fix the E125-A replication failure",
 "diagnosis": "E125-A showed the E124 full arm (85.5%) losing to a compute-matched random search (92.5%) on fresh variants. The failure mode is selection: several depth-2 hypotheses fit the 12 train examples exactly, and fixed-order search keeps a spurious one.",
 "intervention": "Among hypotheses from the SAME enumeration space, fit on the first 8 train rows, select by (fits all 8, accuracy on the held-back 4 train rows, shorter spec). No new hypothesis class, no sealed data used for selection.",
 "splits": {
  "development_variants": [
   7,
   8,
   9
  ],
  "sealed_test_variants": [
   15,
   16,
   17,
   18,
   19
  ],
  "excluded": "0-6 used by E124; 10-14 used by E125-A and therefore not reused as a test set"
 },
 "arms": [
  "baseline_current",
  "compute_matched",
  "cv_filter"
 ],
 "primary_endpoint": "sealed accuracy on variants 15-19 (200 instances) for cv_filter vs baseline_current",
 "success_threshold": "cv_filter beats BOTH baseline_current and compute_matched by >= 5.0 percentage points, exact McNemar p < 0.05 vs baseline_current",
 "failure_rule": "if the gain is below +5.0 pts or p >= 0.05, report NEGATIVE and do not promote",
 "budget": "local CPU only, $0",
 "limits": [
  "constructed generator",
  "same eight paradigms",
  "no external benchmark",
  "no human review"
 ]
}
