# E109 Preregistration

Written before sealed runs: **2026-11-19T15:07:39Z**  
Sealed commitment: `8e7f8d25b3b39d9b86cdb1bb0cb0761c1ddfa3c80f3a7d189b812ecf51fd5977`

## Frozen design

```json
{
 "alpha_label": 0.01,
 "alpha_struct": 0.05,
 "arms": [
  "e109",
  "random",
  "fixed",
  "uncertainty",
  "e108"
 ],
 "bh_family": "frozen hypotheses per environment",
 "budget": 20,
 "min_accuracy": 0.75,
 "min_effect": 0.18,
 "min_interventions": 12,
 "n_obs": 40,
 "permutations": 399,
 "promotion": "best validated rule, one per environment"
}
```

## Gates

- **G1**: clean baseline and historical verifiers pass
- **G10**: zero unauthorized effects, zero after KILL, loopback distinguished from production
- **G11**: standalone verifier, no CAIN imports, hash-pinned raw evidence
- **G12**: all negative results published, no AGI or third-party certification claim
- **G2**: >=6 families, >=4 rule classes, >=2 held-out, no family templates
- **G3**: FDR <= 5% and zero promotions on null/shuffled controls
- **G4**: e109 > strongest matched baseline in >=4 of 6 families
- **G5**: >=3 transfer directions, >=2 positive, negative-transfer detection
- **G6**: lower experiments-to-discovery than random probing at equal budgets
- **G7**: memory suspension/revalidation, poisoning refused, survives restart
- **G8**: calibration and abstention measured, unsupported certainty = 0
- **G9**: collective contributions measured, Byzantine false discovery rejected, no duplicate inflation

No threshold was changed after the sealed set was opened.
