{
 "evolution": "E115 Verified Memory Organism",
 "frozen_at": "2026-11-19",
 "negative_results": "published regardless of outcome",
 "question": "Does evidence-native memory make a model learn across sessions better than standard memory, on an external benchmark, and stay correct under memory attack?",
 "schema": "cain.e115.prereg.v1",
 "trackA": {
  "arms": [
   "M0_none",
   "M1_naive",
   "M2_evidence",
   "M3_causal",
   "M4_validated",
   "R_recent"
  ],
  "benchmark": "AMA-Bench real-world set (HF AMA-bench/AMA-bench @ a5777378, MIT)",
  "budget_tokens": [
   1500,
   4000
  ],
  "floors": [
   "question-echo baseline",
   "M0 no memory"
  ],
  "judge_accuracy": "BLOCKED (LLM judge unavailable: no GPU, no API credit)",
  "model_free_diagnostic": "novel-token recall of retrieved context (dev only; not a benchmark score)",
  "models": {
   "Qwen2.5-1.5B local": "after E113 frees the CPU",
   "claude-haiku-5-5": "BLOCKED until API credit",
   "claude-opus-5-5": "BLOCKED until API credit"
  },
  "pass": "M4 - M1 >= +5 F1 on sealed, 95% CI (paired bootstrap over questions) excludes 0, same model and budget",
  "primary_metric": "official token F1 (AMA-Hub evaluation_metrics.py, verbatim)",
  "secondary": [
   "novel-token F1",
   "exact match"
  ],
  "split_hash": "0027643f7de9e0d34520c45f7bb7169d9eb6e3871e856667161a9da60e6a8506"
 },
 "trackB": {
  "attacks": "memverify.make_attacks rate 0.05 seed 11501 (5 classes)",
  "pass": "CAIN memory (M4) loses <= 1 point; M1 loses >= 5 points; residual (lying_observation) reported separately"
 },
 "trackC": {
  "disclosure": "P1's grammar contains the E114 shift generator's forms; frontier rules are verified inexpressible in P1",
  "pass": "P2 identification on frontier sealed >= 0.30 with false confirmations <= 1/60 (P0/P1 cannot express frontier rules); shift sealed >= 0.60",
  "proposers": {
   "P0_fixed": "E114 grammar",
   "P1_enum": "open enumerator",
   "P2_model": "Claude (BLOCKED: credit) / local Qwen (pending CPU)"
  },
  "settings": {
   "K0": 24,
   "K1": 24,
   "P": 8,
   "ctol": 0.0,
   "reps": 3,
   "tol": 0.15
  },
  "settings_chosen_on": "dev only (val_shift 0-29, frontier dev 0-29); chosen for fewest false confirmations",
  "spec_hashes_sha256": {
   "frontier": "d1fd5558f0baef215e6079b781481e1add976cab454f59fba7f484731b5ee813",
   "shift": "a74a9721d0d8678489b46503d618fdd422ecd113b6b9dd78dae7338f2607a66d"
  },
  "splits": {
   "frontier_sealed": "e115 frontier sealed 0-59",
   "shift_sealed": "E114 envgen shift 0-59"
  }
 }
}
