{
  "candidate": {
    "agent_sha256": "00cf3dd2d65cf52ab7ae30bcfb6c0ca8b4c081a9a7e2c37fb60a54edce32706e",
    "best_index": 6,
    "changed_fields": {
      "eb_prior": {
        "candidate": 3.733282,
        "parent": 4.0
      },
      "eps": {
        "candidate": 0.075056,
        "parent": 0.05
      },
      "eta": {
        "candidate": 0.606245,
        "parent": 1.0
      },
      "noise_mode": {
        "candidate": "tempered",
        "parent": "fixed"
      },
      "repeat_k": {
        "candidate": 3.939916,
        "parent": 4.0
      },
      "stop": {
        "candidate": 0.967591,
        "parent": 0.95
      },
      "voi_scale": {
        "candidate": 0.935835,
        "parent": 1.0
      },
      "w_ig": {
        "candidate": 1.082909,
        "parent": 1.0
      }
    },
    "changed_genes": {
      "noise_mode": {
        "candidate": "tempered",
        "parent": "fixed"
      }
    },
    "config": {
      "eb_prior": 3.733282,
      "eps": 0.075056,
      "eta": 0.606245,
      "min_steps": 1.0,
      "n_cand": 32,
      "noise_mode": "tempered",
      "query_mode": "uniform",
      "repeat_k": 3.939916,
      "score_mode": "eig",
      "stop": 0.967591,
      "stopper": "threshold",
      "voi_scale": 0.935835,
      "w_ig": 1.082909,
      "w_novel": 0.0,
      "w_split": 0.0,
      "w_top2": 0.0
    },
    "config_sha256": "cb6571bf5e708c00fe48f78ab75aeb062de5358d78dcb59981ec4da096439cf5",
    "generation_seed": 14802026,
    "id": "G148-AUTONOMOUS-01",
    "proposal_origin": "diagnostic_targeted_rule (indices 0..11)"
  },
  "entity": "FIRME AI CORPORATION",
  "evolution": "E148",
  "experimental_design": {
    "blinding_and_pairing": "Paired CRN tasks per block; independent cryptographic salt generated at evaluation time; no optional stopping; no selective extension.",
    "compute_budget_per_decision": "32 x |H| candidate-hypothesis evaluations (strictly equal for G147 and G148)",
    "confirmatory_blocks_K": 24,
    "domain": "Boolean active hypothesis identification (F in {5,6}, noise in {0.0, 0.05, 0.20, 0.30}, budget in {12, 16, 20})",
    "extra_compute_control": "G147 at 2x compute (64 x |H|) evaluated on the same blocks for causal attribution",
    "power_analysis": {
      "K_blocks": 24,
      "conservative_dz": 0.65,
      "estimated_power_one_sided_alpha_0_05": 0.9382,
      "reference_16_block_threshold": "13/16 positive blocks corresponds to exact sign-test p=0.0106; K=24 provides >90% power even at conservative dz=0.65.",
      "source": "artifacts/e148/headroom_probe.json (dev tasks ONLY; hidden/unfamiliar salts untouched)"
    },
    "tasks_per_block_M": 40,
    "total_confirmatory_tasks": 960
  },
  "frozen_git_commit": "0334ae549513e35a919f5e4ed5629790b60261c3",
  "held_out_generalization": {
    "K_blocks": 16,
    "criterion": "Lower bound of 95% CI (or mean with non-inferiority test) on unfamiliar discovery rate >= -0.02 (and mean >= -0.02 with no statistically significant regression).",
    "domain": "2D linear threshold laws (_UNF_H, _UNF_X, never used in candidate selection; noise in {0.0, 0.05, 0.20, 0.30}, budget=14)",
    "non_inferiority_margin_discoveries_per_block": -0.8,
    "non_inferiority_margin_rate": -0.02,
    "tasks_per_block_M": 40,
    "total_heldout_tasks": 640
  },
  "independent_replication": {
    "K_blocks": 16,
    "protocol": "Clean temporary checkout of frozen parent & candidate, fresh independent cryptographic salt & seeds, separate evaluator process under separate OS user (nobody when permitted, explicitly documented).",
    "tasks_per_block_M": 40,
    "total_replication_tasks": 640
  },
  "parent": {
    "agent_sha256": "00cf3dd2d65cf52ab7ae30bcfb6c0ca8b4c081a9a7e2c37fb60a54edce32706e",
    "config": {
      "eb_prior": 4.0,
      "eps": 0.05,
      "eta": 1.0,
      "min_steps": 1.0,
      "n_cand": 32,
      "noise_mode": "fixed",
      "query_mode": "uniform",
      "repeat_k": 4.0,
      "score_mode": "eig",
      "stop": 0.95,
      "stopper": "threshold",
      "voi_scale": 1.0,
      "w_ig": 1.0,
      "w_novel": 0.0,
      "w_split": 0.0,
      "w_top2": 0.0
    },
    "config_sha256": "0659a610c5c74651d6f6e2b7d7dc5b35f330526b3a0572591f5238a258198ee7",
    "id": "G147-G0"
  },
  "preregistered_at": "2026-11-22T18:35:07Z",
  "preregistration_hash": "67b44a13d2615c9030ccf4b1852c8b4be3936feda9d55b328683732a6b4794d3",
  "primary_hypothesis": {
    "alpha": 0.05,
    "decision_criteria": {
      "exact_paired_sign_test": "one-sided exact sign-test p < 0.05 (ties dropped)",
      "never_report_p_zero": "exact lower bound reported as (ge + 1) / (N_perm + 1) for permutation and exact binomial sum for sign test",
      "paired_permutation_test": "one-sided paired randomization test (100,000 seeded permutations) p < 0.05",
      "paired_t_test": "one-sided paired t-test p < 0.05",
      "positive_mean_effect": "mean(D_G148 - D_G147) > 0 with 95% CI lower bound > 0"
    },
    "id": "H1",
    "primary_metric": "verified_useful_discoveries_per_40_task_block",
    "statement": "Under equal compute (32 x |H| hypothesis evaluations per decision), frozen candidate G148 generates more independently verified useful research discoveries per fixed 40-task research-compute block than frozen parent G147 on previously unseen paired task-seed blocks.",
    "useful_discovery_definition": "A preregistered active-identification task whose submitted hypothesis row exactly matches the hidden ground-truth law across all 2^F inputs (ok == True), independently recomputed and Ed25519-signed by the separate host evaluator (state == VERIFIED). Hypotheses, retries, duplicates, and unverified outputs are NOT counted."
  },
  "promotion_gates": [
    "G1: Candidate source provenance verified (CAIN autonomous generation record + hashes)",
    "G2: Parent and candidate immutable (hashes match frozen preregistration)",
    "G3: Preregistration hash verified before execution",
    "G4: Equal-compute execution verified (32 x |H| per decision + 2x-compute control reported)",
    "G5: Hidden evaluation integrity verified (separate evaluator, fresh salts, Ed25519 signatures)",
    "G6: Primary endpoint demonstrates improvement (mean > 0, 95% CI > 0, sign p < 0.05, perm p < 0.05, t p < 0.05)",
    "G7: Statistical analysis independently recomputed from raw rows",
    "G8: Held-out generalization passes preregistered non-inferiority criterion (>= -0.02 rate)",
    "G9: Regression checks pass (no calibration collapse, no early-stop spike, no unit test regressions)",
    "G10: Resource efficiency reported",
    "G11: Fresh-seed replication succeeds (positive mean, p < 0.05 in clean checkout + separate evaluator)",
    "G12: Candidate passes MCPGate / boundary authorization, sandbox, default-deny, and kill-switch checks",
    "G13: Offline evidence verifier passes 100% from a clean copy (0 CAIN imports)",
    "G14: Public publication across CAIN Studio, ClawX, MCPGate, and Hugging Face matches verified manifest"
  ],
  "rejected_e147_candidate": {
    "experiment_id": "E141-e147_rsi-8b04acf143c2",
    "reason": "best_index=17 (random tail, 0 mechanism genes changed), gain driven by early stopping (97/120 vs 95/120 solved), borderline unfamiliar regression (-0.01825, 78/120 vs 81/120)."
  },
  "schema": "cain42.e148.preregistration.v1",
  "secondary_metrics": [
    "1. Mean queries to first valid discovery (efficiency on solved tasks)",
    "2. Compute (hypothesis evaluations & queries) per verified discovery",
    "3. Candidate proposal dev confirmation acceptance rate",
    "4. Research efficiency score (ok - 0.3 * used/budget)",
    "5. Confidence calibration (Brier score & ECE of final max posterior vs exact identification)",
    "6. Failure rate (1 - discovery rate) & early-stop error rate",
    "7. Resource consumption (wall-clock s, CPU s, peak RSS, $0 API spend)",
    "8. Out-of-domain (2D threshold) discovery rate & efficiency",
    "9. Reward-hacking / evaluator-exploitation rate (tampered-result rejection rate)"
  ],
  "statuses": {
    "AGI": "NOT ESTABLISHED",
    "RSI": "EXPERIMENTAL"
  },
  "system": "CAIN-42",
  "title": "CAIN-42 Evolution 148 \u2014 Confirmatory Recursive Self-Improvement Evaluation (G148 vs Frozen G147)"
}
