{
  "ten_role_scientific_discovery": {
    "schema": "cain.e140.scientific_discovery.v1",
    "generated_at": "2026-11-21T23:50:33Z",
    "duration_s": 17.85,
    "rsi_category": "2_workflow_and_research_process_evolution + 4_model_assisted_candidate_generation",
    "roles_10_separated": {
      "model_backed_roles": [
        "03_hypothesis_generator",
        "04_experiment_designer",
        "07_counter_hypothesis_critic"
      ],
      "deterministic_roles": [
        "01_literature_researcher",
        "02_evidence_extractor",
        "05_statistical_analyst",
        "06_experiment_executor",
        "08_reproduction_agent",
        "09_safety_governance_reviewer",
        "10_independent_evidence_verifier"
      ]
    },
    "live_ollama_roles": {
      "model_id": "qwen2.5:0.5b",
      "calls": 3,
      "tokens": 319,
      "paid_usd": 0.0,
      "trace": [
        {
          "role": "03_hypothesis_generator",
          "ok": true,
          "latency_ms": 8446.5,
          "snippet": "{   \"hypothesis\": \"Observational zero-intervention residuals reveal differences between gr"
        },
        {
          "role": "04_experiment_designer",
          "ok": true,
          "latency_ms": 5178.7,
          "snippet": "{   \"hypothesis\": \"Observational zero-intervention residuals can distinguish between bimod"
        },
        {
          "role": "07_counter_hypothesis_critic",
          "ok": true,
          "latency_ms": 4205.7,
          "snippet": "{   \"hypothesis\": \"Observational zero-intervention residuals reveal differences between gr"
        }
      ]
    },
    "total_evaluations": 500,
    "cain_accuracy": 0.916,
    "baseline_accuracy": 0.544,
    "comparison_vs_baseline": {
      "test": "exact_mcnemar_arbitrary_precision",
      "n": 500,
      "rate_a": 0.916,
      "rate_b": 0.544,
      "gain": 0.372,
      "a_only": 211,
      "b_only": 25,
      "p_value": 7.447806735706712e-38,
      "log10_p": -37.128,
      "p_display": "7.448e-38",
      "underflowed_float64": false,
      "effect_size": 0.8946,
      "effect_size_type": "cohens_h",
      "confidence_interval_diff": [
        0.3214,
        0.4226
      ]
    },
    "citation_correctness_rate": 1.0,
    "unsupported_claim_rate": 0.0,
    "verifier_vetoes_exercised": 1,
    "replication_success_rate": 1.0,
    "cycles": [
      {
        "id": "SCI-E140-01",
        "lit": "LIT-08-CAUSAL",
        "claim": "Negative-tail + 2-means variance reduction separates binary vs continuous confounders",
        "verifier_action": "ACCEPTED_REPLICATED"
      },
      {
        "id": "SCI-E140-02",
        "lit": "LIT-11-ARC-AGI-3",
        "claim": "Broad+Deep exploration beats single-mode exploration on hidden transition families",
        "verifier_action": "ACCEPTED_REPLICATED"
      },
      {
        "id": "SCI-E140-03",
        "lit": "LIT-12-RSIAGENT",
        "claim": "2-cycle meta-proposer improvement transfers to independent unseen task family",
        "verifier_action": "ACCEPTED_REPLICATED"
      },
      {
        "id": "SCI-E140-04",
        "lit": "LIT-07-CONTINUAL",
        "claim": "Contradiction-detecting expiring memory cuts stale-memory harm under distribution shift",
        "verifier_action": "ACCEPTED_REPLICATED"
      },
      {
        "id": "SCI-E140-05",
        "lit": "LIT-10-REWARDHACK",
        "claim": "Overfit proxy-maximizing proposer generalizes without sealed validation",
        "verifier_action": "REJECTED_BY_INDEPENDENT_VERIFIER"
      }
    ],
    "computational_simulation_disclosure": "All E140 discovery experiments are computational/algorithmic; zero wet-lab or physical-world experimental discoveries are claimed.",
    "scientific_verdict": "PASS",
    "digest": "6e240e3f3e1886708a6961d48d3dfc72063e2dab6156940bb7d2aa285b375879"
  }
}