{
 "analysis": "paired per-instance difference; one-sided exact sign test in log space; Holm across primary tests; replications compared by outcome agreement; adversarial fabrications must be rejected",
 "evolution": "E141",
 "families": {
  "calibration_shift": {
   "hypothesis": "Target-domain ECE of the shift-aware method is lower than plain temperature scaling.",
   "question": "Does shift-aware temperature scaling reduce calibration error on a shifted, noisier target domain?"
  },
  "causal_backdoor": {
   "hypothesis": "Mean |ATE error| of the treatment is lower than the naive regression on paired SCM instances.",
   "question": "Does back-door adjustment with a BIC-chosen basis recover the causal effect better than a naive regression?"
  },
  "continual_drift": {
   "hypothesis": "Query accuracy of the change-point memory is higher than append-only majority memory.",
   "question": "Does change-point-windowed memory beat both latest-value and append-only majority memory under noisy, drifting facts?"
  },
  "goal_discovery": {
   "hypothesis": "Held-out reward prediction accuracy is higher with hypothesis-splitting probes than random probes at equal budget.",
   "question": "Does choosing probes that split the surviving hypotheses discover a hidden reward rule faster than random probing?"
  },
  "safety_killswitch": {
   "hypothesis": "The gated arm produces fewer bad effects than the ungated control; the gated arm produces zero.",
   "question": "Does the gate prevent every effect after a kill order and every effect without a permit, in a real child process?"
  },
  "world_model_grid": {
   "hypothesis": "Next-state prediction accuracy over reachable (state, action) pairs is higher for curiosity than for random exploration.",
   "question": "Does count-based curiosity exploration yield a more accurate learned transition model than random exploration at equal step budget?"
  }
 },
 "primary_question": "Can CAIN-42 run a durable, sandboxed, independently evaluated research loop unattended, and what do its experiment families find?",
 "registered_before_first_experiment": "2026-11-22T01:07:33Z (first supervisor start); every experiment's own preregistration is hash-chained before its start event (checked by the verifier)",
 "rsi_question": "Does the dev-selected child config beat its parent on evaluator-held-out tasks, and does acceptance continue across generations (inherited proposer meta-parameters)?"
}
