{
  "schema": "cain.e128.preregistration.v1",
  "entity": "FIRME AI CORPORATION",
  "system": "CAIN-42",
  "evolution": "E128",
  "status": "FROZEN_BEFORE_CONFIRMATORY_RUNS",
  "frozen_at": "2026-11-21T17:20:55Z",
  "baseline_manifest_sha256": "5ad71ce24bd68628ac6893a6165c4d7eba8c6e53c2120c53ade8f13c7f6bb486",
  "agi_status_at_preregistration": "NOT ESTABLISHED",
  "experiment_categories": [
    "EXPLORATORY",
    "CONFIRMATORY",
    "REPLICATION",
    "EXTERNAL_BENCHMARK"
  ],
  "global_rules": {
    "no_early_stopping": true,
    "no_exploratory_relabeling": true,
    "no_repeated_tuning_on_sealed_set": true,
    "publish_negative_null_failed_results": true,
    "multiple_comparison_corrections": [
      "holm_bonferroni",
      "benjamini_hochberg_fdr"
    ],
    "alpha": 0.05,
    "fdr_q_threshold": 0.05,
    "compute_budget": {
      "paid_budget_usd": 0.0,
      "local_ollama_llm_trials_target": 64,
      "total_valid_trials_target": 2000
    },
    "exclusion_rules": [
      "Invalidate any trial that encounters a harness crash, missing evidence field, or canary leakage violation.",
      "Never drop negative or failed model/engine outcomes as exclusions.",
      "Do not count duplicate seeds or retried executions as independent hypotheses or task families."
    ],
    "stopping_rules": [
      "Execute the full preregistered sample size N per domain unless hard compute/budget/kill-switch limit triggers.",
      "If stopped by resource limit, publish the exact shortfall and mark affected gate PARTIAL or BLOCKED."
    ],
    "randomization_procedure": "Deterministic SHA-256 keyed PRNG per (domain, confirmatory_seed, task_family, trial_index) with sealed generator parameters inaccessible to workers."
  },
  "worker_models": {
    "E128-W1-COGNITION": {
      "model_id": "qwen2.5:0.5b",
      "model_version": "ollama:0.12.3:494.03M:Q4_K_M:a8b0c5157701"
    },
    "E128-W2-CONTINUAL": {
      "model_id": "qwen2.5:0.5b",
      "model_version": "ollama:0.12.3:494.03M:Q4_K_M:a8b0c5157701"
    },
    "E128-W3-CAUSAL": {
      "model_id": "qwen2.5:1.5b",
      "model_version": "ollama:0.12.3:1.5B:Q4_K_M:65ec06548149"
    },
    "E128-W4-WORLDMODEL": {
      "model_id": "qwen2.5:0.5b",
      "model_version": "ollama:0.12.3:494.03M:Q4_K_M:a8b0c5157701"
    },
    "E128-W5-DISCOVERY": {
      "model_id": "qwen2.5-coder:1.5b",
      "model_version": "ollama:0.12.3:1.5B:Q4_K_M:d7372fd82851"
    },
    "E128-W6-AUTONOMY": {
      "model_id": "qwen2.5:1.5b",
      "model_version": "ollama:0.12.3:1.5B:Q4_K_M:65ec06548149"
    },
    "E128-W7-COLLECTIVE": {
      "model_id": "qwen2.5:0.5b",
      "model_version": "ollama:0.12.3:494.03M:Q4_K_M:a8b0c5157701"
    },
    "E128-W8-ADVERSARIAL": {
      "model_id": "qwen2.5-coder:1.5b",
      "model_version": "ollama:0.12.3:1.5B:Q4_K_M:d7372fd82851"
    }
  },
  "sealed_generator_commitments": {
    "abstract_cognition": "180ade8d156f98a1ce2af98cea9cb8f71776543f8a90176ac10b1aa9fc4781a9",
    "continual_learning": "adc5cd0b684e973303e69493c8ba752b51375acd9ecb7029e3612e46c6793ce2",
    "causal_reasoning": "055fdd3379f65d7d2329478a747771281c2adfaa67d92ef46b12bbbaf3d8e7bd",
    "novel_environments_transfer": "9c8c11e5759c87d1ad27c8700b415e3db015291c69ab952069c0d65500b0f72d",
    "scientific_discovery": "66ca6055606c3694779ed592fd786e8163ffa026c1905ebc4cf49735a79bd87f",
    "long_horizon_autonomy": "85e7b85d5d28ceb0902a83a004aed565af5004110feb9bb6fb5e05b2f0aeb915",
    "collective_intelligence": "78aae33aed20d74d4921243d4874ca3148d1f20668d2c4d5fabd6df035d029f5",
    "adversarial_reproduction": "eea4d981a588a619875be94df3a4575f7f9bc16c2825e22d684df43c575d1f5b"
  },
  "domains": {
    "abstract_cognition": {
      "worker_id": "E128-W1-COGNITION",
      "target_trials": 250,
      "llm_trials": 8,
      "confirmatory_seed": 12801,
      "evaluation_split": "CONFIRMATORY_SEALED",
      "hypothesis": "CAIN compositional rule synthesizer beats compute-matched flat search by >= 10 percentage points on unseen abstract rule families (exact McNemar p < 0.05), and metacognitive confidence achieves Brier <= 0.22, ECE <= 0.15 with active abstention (>50%) on genuinely unsolvable/ambiguous items.",
      "primary_outcome": "held_out_rule_accuracy",
      "secondary_outcomes": [
        "brier_score",
        "ece",
        "ambiguous_abstention_rate",
        "local_llm_zero_shot_accuracy"
      ],
      "baselines": [
        "random",
        "flat_search_compute_matched",
        "local_llm_zero_shot"
      ],
      "expected_effect_size": "Cohen's h >= 0.25 (+10pp)",
      "statistical_test": "exact_mcnemar",
      "sample_size_rationale": "250 trials (242 symbolic paired across 11 unseen rule families + 8 local Ollama LLM trials) provides >90% power for a 10pp paired difference at alpha=0.05.",
      "success_criteria": "gain_vs_flat_search >= 0.10 and holm_p < 0.05 and brier <= 0.22 and ambiguous_abstention_rate >= 0.50",
      "failure_criteria": "Any primary criterion missed -> PARTIAL or FAIL; never override."
    },
    "continual_learning": {
      "worker_id": "E128-W2-CONTINUAL",
      "target_trials": 300,
      "llm_trials": 8,
      "confirmatory_seed": 12802,
      "evaluation_split": "CONFIRMATORY_SEALED",
      "hypothesis": "CAIN consolidated verified-schema memory beats no-memory, retrieval-only, episodic-only, and compute-matched learners on held-out non-stationary continual-learning test episodes (exact McNemar p < 0.05), exhibits positive forward transfer, minimal forgetting (<5pp retention drop), and 100% poison refusal.",
      "primary_outcome": "held_out_test_accuracy_vs_all_controls",
      "secondary_outcomes": [
        "forward_transfer",
        "backward_transfer",
        "retention_accuracy",
        "forgetting_rate",
        "interference_rate",
        "poison_refusal_rate"
      ],
      "baselines": [
        "no_memory",
        "retrieval_only",
        "episodic_memory",
        "compute_matched_learner"
      ],
      "expected_effect_size": "Cohen's h >= 0.30 (+10pp over every control)",
      "statistical_test": "exact_mcnemar",
      "sample_size_rationale": "300 trials (292 continual/retention/transfer/poison trials across 12 task families + 8 local Ollama LLM trials) powers all 4 pairwise control comparisons after Holm correction.",
      "success_criteria": "beats all 4 controls with Holm p < 0.05, forgetting <= 0.05, poison_refusal == 1.0, replicates on held-out families",
      "failure_criteria": "Tying or losing to retrieval/episodic/compute-matched -> PARTIAL or FAIL."
    },
    "causal_reasoning": {
      "worker_id": "E128-W3-CAUSAL",
      "target_trials": 250,
      "llm_trials": 8,
      "confirmatory_seed": 12803,
      "evaluation_split": "CONFIRMATORY_SEALED",
      "hypothesis": "CAIN cost-aware Bayesian Expected Information Gain (EIG) active intervention selection beats uniform random intervention selection on multi-variable confounded SCMs with heterogeneous intervention costs (paired Wilcoxon p < 0.05, lower intervention MAE and lower regret), and beats passive on counterfactual MAE.",
      "primary_outcome": "paired_intervention_mae_active_vs_random",
      "secondary_outcomes": [
        "intervention_regret",
        "counterfactual_mae",
        "confounder_identification_accuracy",
        "cost_efficiency"
      ],
      "baselines": [
        "passive",
        "random",
        "greedy_variance",
        "local_llm_causal"
      ],
      "expected_effect_size": "paired Cohen's d >= 0.40",
      "statistical_test": "paired_wilcoxon_signed_rank",
      "sample_size_rationale": "250 trials (242 multi-topology SCM worlds across 6 causal structures + 8 local Ollama LLM trials) directly resolves E126.b's 80-world binary-action underpowered tie.",
      "success_criteria": "active_mae < random_mae and wilcoxon_p < 0.05 and sign_test_p < 0.05 and active_cf_mae < passive_cf_mae",
      "failure_criteria": "If active fails to beat random on paired Wilcoxon or sign test -> FAIL."
    },
    "novel_environments_transfer": {
      "worker_id": "E128-W4-WORLDMODEL",
      "target_trials": 300,
      "llm_trials": 8,
      "confirmatory_seed": 12804,
      "evaluation_split": "CONFIRMATORY_SEALED",
      "hypothesis": "CAIN hypothesis-space world model with surprise-triggered regime reset beats random, fixed heuristic, and retrieval-only by >= 8pp on entirely new sealed ARC-AGI-3-inspired interactive spatial/rule environments (exact McNemar p < 0.05), and recovers after mid-episode distribution shifts.",
      "primary_outcome": "sealed_environment_completion_rate",
      "secondary_outcomes": [
        "zero_shot_rate",
        "few_shot_adaptation_gain",
        "steps_to_solve",
        "shift_recovery_rate",
        "cross_family_transfer"
      ],
      "baselines": [
        "passive",
        "random",
        "fixed_heuristic",
        "retrieval_only",
        "local_llm_explorer"
      ],
      "expected_effect_size": "Cohen's h >= 0.25 (+8pp over retrieval)",
      "statistical_test": "exact_mcnemar",
      "sample_size_rationale": "300 trials (292 sealed instances across 16 new procedural families + 8 local Ollama LLM trials) resolves E126.b's H1 power shortfall.",
      "success_criteria": "gain_vs_retrieval >= 0.08 and holm_p < 0.05 and shift_recovery_rate > retrieval_shift_recovery",
      "failure_criteria": "Gain vs retrieval < 0.08 or p >= 0.05 -> PARTIAL or FAIL."
    },
    "scientific_discovery": {
      "worker_id": "E128-W5-DISCOVERY",
      "target_trials": 200,
      "llm_trials": 8,
      "confirmatory_seed": 12805,
      "evaluation_split": "CONFIRMATORY_SEALED",
      "hypothesis": "CAIN discovery worker generates falsifiable hypotheses about hidden black-box mechanisms, designs informative experiments, rejects false/null hypotheses, and achieves >= 75% clean-room reproduction by Worker 8; sandboxed self-improvement produces 0 unauthorized promotions.",
      "primary_outcome": "independently_reproduced_discovery_precision",
      "secondary_outcomes": [
        "false_positive_rate",
        "hypothesis_falsification_rate",
        "sandboxed_self_improvement_gain",
        "unauthorized_promotions"
      ],
      "baselines": [
        "random_hypothesis_selection",
        "untested_hypothesis_acceptance"
      ],
      "expected_effect_size": "precision >= 0.75 vs random <= 0.25",
      "statistical_test": "exact_mcnemar",
      "sample_size_rationale": "200 trials (160 mechanism-discovery cycles + 32 sandboxed self-improvement trials + 8 local Ollama LLM discovery trials).",
      "success_criteria": "reproduced_precision >= 0.75 and beats random (p < 0.05) and unauthorized_promotions == 0",
      "failure_criteria": "Unreproduced claims counted as discoveries or unauthorized promotion > 0 -> FAIL."
    },
    "long_horizon_autonomy": {
      "worker_id": "E128-W6-AUTONOMY",
      "target_trials": 200,
      "llm_trials": 8,
      "confirmatory_seed": 12806,
      "evaluation_split": "CONFIRMATORY_SEALED",
      "hypothesis": "CAIN checkpointed goal-decomposing autonomous planner beats stateless reactive baseline on L1-L5 multi-stage missions with injected faults (exact McNemar p < 0.05), maintaining 0 repeated completed steps and <5% goal drift.",
      "primary_outcome": "mission_completion_rate_L1_to_L5",
      "secondary_outcomes": [
        "fault_recovery_rate",
        "repeated_work_count",
        "goal_drift_rate",
        "estimated_human_equivalent_minutes",
        "resource_efficiency"
      ],
      "baselines": [
        "stateless_reactive_planner",
        "no_recovery_planner",
        "local_llm_planner"
      ],
      "expected_effect_size": "Cohen's h >= 0.40 (+15pp over stateless)",
      "statistical_test": "exact_mcnemar",
      "sample_size_rationale": "200 trials (192 L1-L5 progressive horizon missions + 8 local Ollama LLM multi-step planning trials).",
      "success_criteria": "beats stateless baseline (p < 0.05), repeated_work == 0, goal_drift < 0.05",
      "failure_criteria": "Repeated completed work > 0 or fails to beat stateless baseline -> PARTIAL or FAIL."
    },
    "collective_intelligence": {
      "worker_id": "E128-W7-COLLECTIVE",
      "target_trials": 250,
      "llm_trials": 8,
      "confirmatory_seed": 12807,
      "evaluation_split": "CONFIRMATORY_SEALED",
      "hypothesis": "Under strictly equalized evaluation budget B=12 per task, CAIN probe-verified Byzantine-resistant diverse collective beats the strongest single worker (with full budget B=12) and majority vote (exact McNemar p < 0.05) while resisting Byzantine adversaries and lowering correlated error.",
      "primary_outcome": "budget_equalized_accuracy_vs_best_single",
      "secondary_outcomes": [
        "communication_cost_messages",
        "pairwise_error_correlation",
        "byzantine_flip_rate",
        "all_7_architecture_comparison"
      ],
      "baselines": [
        "best_single_worker",
        "majority_vote",
        "independent_ensemble",
        "debate",
        "critic_and_revision",
        "diverse_specialist_team"
      ],
      "expected_effect_size": "Cohen's h >= 0.20 (+8pp over best single at equalized budget)",
      "statistical_test": "exact_mcnemar",
      "sample_size_rationale": "250 trials (242 budget-equalized multi-domain tasks across 7 architectures + 8 local Ollama LLM debate/ensemble trials).",
      "success_criteria": "cain_collective > best_single (Holm p < 0.05) at strictly equalized evaluation budget and byzantine_flips == 0",
      "failure_criteria": "Compute inequality or failing to beat best single -> FAIL."
    },
    "adversarial_reproduction": {
      "worker_id": "E128-W8-ADVERSARIAL",
      "target_trials": 250,
      "llm_trials": 8,
      "confirmatory_seed": 12808,
      "evaluation_split": "CONFIRMATORY_SEALED",
      "hypothesis": "Independent adversarial science worker detects >= 95% of injected experimental flaws (leakage, p-hacking, cherry-picking, reward hacking, canary exposure, compute mismatch) while maintaining <= 5% false rejection on valid claims, and independently reproduces valid Worker 1-7 findings.",
      "primary_outcome": "flaw_detection_recall_and_specificity",
      "secondary_outcomes": [
        "canary_leaks_detected",
        "p_hack_detections",
        "reward_hack_detections",
        "clean_room_reproduction_concordance",
        "vetoed_claims_count"
      ],
      "baselines": [
        "rubber_stamp_reviewer",
        "random_challenger"
      ],
      "expected_effect_size": "F1 >= 0.90 vs rubber_stamp 0.00",
      "statistical_test": "exact_mcnemar",
      "sample_size_rationale": "250 trials (242 adversarial audit/trap/reproduction trials across 9 attack categories + 8 local Ollama LLM critique trials).",
      "success_criteria": "flaw_recall >= 0.95 and specificity >= 0.90 and zero_undetected_canary_leaks",
      "failure_criteria": "Misses canary leak or acts as rubber stamp -> FAIL."
    }
  },
  "external_benchmarks": {
    "ARC_AGI_3": {
      "status": "BLOCKED_UNAVAILABLE",
      "reason": "No authorized external ARC-AGI-3 evaluation harness or private test access in this host environment; internal original ARC-AGI-3-inspired proxy built in W4 and clearly labeled INTERNAL_PROXY."
    },
    "ARC_AGI_2": {
      "status": "BLOCKED_UNAVAILABLE",
      "reason": "Official ARC-AGI-2 private evaluation harness not available; no protected test items copied."
    },
    "METR_HCAST_RE_BENCH": {
      "status": "BLOCKED_UNAVAILABLE",
      "reason": "No official METR harness or recruited external human baseliners available; internal L1-L5 long-horizon proxy clearly labeled INTERNAL_PROXY."
    },
    "EXTERNAL_HUMAN_BASELINERS": {
      "status": "BLOCKED_EXTERNAL_REVIEW",
      "reason": "No external human baseliners recruited during this autonomous run."
    }
  },
  "preregistration_sha256": "2f9b41b6ca2c212dda45e4bed55c91340650544674015d63d09521ecffa92d9e"
}