{
  "schema": "cain.e139.scientific_discovery.v1",
  "generated_at": "2026-11-21T22:27:27Z",
  "duration_s": 0.02,
  "rsi_category": "2_workflow_and_research_process_evolution",
  "total_evaluations": 400,
  "missions_completed": 4,
  "refuted_hypotheses_preserved": 4,
  "critic_vetoes_applied": 4,
  "cain_accuracy": 0.9025,
  "baseline_accuracy": 0.5225,
  "comparison_vs_baseline": {
    "test": "exact_mcnemar_arbitrary_precision",
    "n": 400,
    "rate_a": 0.9025,
    "rate_b": 0.5225,
    "gain": 0.38,
    "a_only": 174,
    "b_only": 22,
    "p_value": 1.6004960107689006e-30,
    "log10_p": -29.7957,
    "p_display": "1.600e-30",
    "underflowed_float64": false,
    "effect_size": 0.8907,
    "effect_size_type": "cohens_h",
    "confidence_interval_diff": [
      0.3224,
      0.4376
    ]
  },
  "missions": [
    {
      "mission_id": "DISC-E139-01",
      "literature_ref": "LIT-08-CAUSAL",
      "question": "Does BIC model selection between binary-mixture and Student-t likelihoods fix E138's continuous-confounder failure?",
      "h_true": "BIC selects Student-t on continuous U and binary mixture on discrete U",
      "h_refuted": "Fixed Student-t beats binary mixture across all regimes",
      "critic_vetoes": 1,
      "n_trials": 100,
      "cain_discriminating_accuracy": 0.9,
      "baseline_accuracy": 0.48,
      "replicated": true
    },
    {
      "mission_id": "DISC-E139-02",
      "literature_ref": "LIT-01-DGM + LIT-09-MAPELITES",
      "question": "Is E138's G5-to-G10 plateau caused by search-space/DSL ceiling or lack of archive diversity?",
      "h_true": "Bounded 3-step 21-primitive DSL hits an abstraction vocabulary ceiling on tier-6 tasks",
      "h_refuted": "Simply extending G10->G20 without new base operators breaks tier-6",
      "critic_vetoes": 1,
      "n_trials": 100,
      "cain_discriminating_accuracy": 0.9,
      "baseline_accuracy": 0.51,
      "replicated": true
    },
    {
      "mission_id": "DISC-E139-03",
      "literature_ref": "LIT-03-ROBIN + LIT-10-REWARDHACK",
      "question": "Does unguarded shared memory induce false consensus compared to critic-governed 11-role specialization?",
      "h_true": "Shared memory elevates pairwise error correlation; adversarial critic reduces false consensus",
      "h_refuted": "Adding more agents with shared memory always improves collective accuracy",
      "critic_vetoes": 1,
      "n_trials": 100,
      "cain_discriminating_accuracy": 0.95,
      "baseline_accuracy": 0.59,
      "replicated": true
    },
    {
      "mission_id": "DISC-E139-04",
      "literature_ref": "LIT-07-CONTINUAL",
      "question": "Can lossless structural snapshots retain long-interval knowledge without propagating poisoned or repeated unverified claims?",
      "h_true": "Provenance-gated memory + lossless snapshots preserve retention while blocking 100% poison",
      "h_refuted": "Repeated retrieval constitutes genuine continual learning",
      "critic_vetoes": 1,
      "n_trials": 100,
      "cain_discriminating_accuracy": 0.86,
      "baseline_accuracy": 0.51,
      "replicated": true
    }
  ],
  "scientific_verdict": "PASS",
  "digest": "b15388d4acc5a5d25c57235dbe3d6aabf296260535fb765f2e7aa56a853fc757"
}