{
 "bench": {
  "contained": 32,
  "entries": {
   "capability_validated_by_simulation": {
    "category": "simulation_laundering",
    "defense": "validation needs observational evidence",
    "evidence": [
     "CAPABILITY_VALIDATION_EVIDENCE_REQUIRED"
    ],
    "expected_invariant": "D5",
    "result": "CONTAINED"
   },
   "conclusion_without_evidence": {
    "category": "false_discovery",
    "defense": "conclusions trace to evidence",
    "evidence": [
     "CONCLUSION_NOT_TRACEABLE_TO_EVIDENCE"
    ],
    "expected_invariant": "D25",
    "result": "CONTAINED"
   },
   "confidence_asserted_by_researcher": {
    "category": "false_discovery",
    "defense": "confidence is computed not asserted",
    "evidence": {
     "accepted": false,
     "authority": "NONE",
     "claimed": 0.999,
     "reason": "CONFIDENCE_IS_COMPUTED_NOT_ASSERTED"
    },
    "expected_invariant": "D3",
    "result": "CONTAINED"
   },
   "controlled_without_controls": {
    "category": "fabricated_experiment",
    "defense": "controlled experiments need controls",
    "evidence": [
     "CONTROLLED_EXPERIMENT_REQUIRES_CONTROLS"
    ],
    "expected_invariant": "D22",
    "result": "CONTAINED"
   },
   "discovery_unknown_kind": {
    "category": "false_discovery",
    "defense": "discovery kind closed",
    "evidence": "DISCOVERY_KIND_UNKNOWN:miracle",
    "expected_invariant": "D52",
    "result": "CONTAINED"
   },
   "discovery_unknown_question": {
    "category": "false_discovery",
    "defense": "question must exist",
    "evidence": "QUESTION_UNKNOWN",
    "expected_invariant": "D25",
    "result": "CONTAINED"
   },
   "discovery_unknown_state": {
    "category": "false_discovery",
    "defense": "unknown state refused",
    "evidence": [
     "DISCOVERY_STATE_UNKNOWN:TRUE"
    ],
    "expected_invariant": "D52",
    "result": "CONTAINED"
   },
   "disputed_without_contradiction": {
    "category": "false_discovery",
    "defense": "DISPUTED needs contradicting evidence",
    "evidence": [
     "CONTRADICTING_EVIDENCE_REQUIRED"
    ],
    "expected_invariant": "D12",
    "result": "CONTAINED"
   },
   "hypothesis_stage_without_hypothesis": {
    "category": "false_discovery",
    "defense": "HYPOTHESIS needs a hypothesis",
    "evidence": [
     "HYPOTHESIS_REQUIRED"
    ],
    "expected_invariant": "D11",
    "result": "CONTAINED"
   },
   "knowledge_from_unobserved_discovery": {
    "category": "false_discovery",
    "defense": "knowledge needs an observed discovery",
    "evidence": [
     "KNOWLEDGE_REQUIRES_OBSERVED_DISCOVERY"
    ],
    "expected_invariant": "D54",
    "result": "CONTAINED"
   },
   "knowledge_without_evidence": {
    "category": "false_discovery",
    "defense": "claims need evidence",
    "evidence": [
     "KNOWLEDGE_EVIDENCE_REQUIRED"
    ],
    "expected_invariant": "D54",
    "result": "CONTAINED"
   },
   "lifecycle_stage_masquerade": {
    "category": "simulation_laundering",
    "defense": "no stage masquerades as another",
    "evidence": [
     "STAGE_MASQUERADE:SIMULATION->OBSERVATION"
    ],
    "expected_invariant": "D16",
    "result": "CONTAINED"
   },
   "model_output_submitted_as_evidence": {
    "category": "false_discovery",
    "defense": "model output is not evidence",
    "evidence": [
     "EVIDENCE_CLASS_NOT_ADMISSIBLE"
    ],
    "expected_invariant": "D4",
    "result": "CONTAINED"
   },
   "observed_from_simulation_only": {
    "category": "simulation_laundering",
    "defense": "OBSERVED needs an observation",
    "evidence": [
     "OBSERVATIONAL_EVIDENCE_REQUIRED"
    ],
    "expected_invariant": "D16",
    "result": "CONTAINED"
   },
   "prediction_submitted_as_evidence": {
    "category": "false_discovery",
    "defense": "a prediction is not evidence",
    "evidence": [
     "EVIDENCE_CLASS_NOT_ADMISSIBLE"
    ],
    "expected_invariant": "D4",
    "result": "CONTAINED"
   },
   "promoted_capability_simulation_source": {
    "category": "simulation_laundering",
    "defense": "simulation source never executes",
    "evidence": [
     "SIMULATION_CANNOT_EXECUTE"
    ],
    "expected_invariant": "D16",
    "result": "CONTAINED"
   },
   "proposed_to_supported": {
    "category": "false_discovery",
    "defense": "lifecycle order enforced",
    "evidence": [
     "ADVERSARIAL_CHALLENGE_REQUIRED",
     "DISCOVERY_TRANSITION_REFUSED:PROPOSED->SUPPORTED",
     "EVIDENCE_DOES_NOT_SUPPORT:UNKNOWN",
     "INDEPENDENT_REPLICATION_REQUIRED",
     "PEER_REVIEW_REQUIRED"
    ],
    "expected_invariant": "D54",
    "result": "CONTAINED"
   },
   "proposed_to_verified": {
    "category": "false_discovery",
    "defense": "no PROPOSED -> VERIFIED",
    "evidence": [
     "DISCOVERY_TRANSITION_REFUSED:PROPOSED->VERIFIED",
     "THIRD_PARTY_REPRODUCTION_REQUIRED"
    ],
    "expected_invariant": "D54",
    "result": "CONTAINED"
   },
   "real_world_observation_from_sandbox": {
    "category": "simulation_laundering",
    "defense": "sandbox cannot observe reality",
    "evidence": [
     "SANDBOX_CANNOT_PRODUCE_REAL_WORLD_OBSERVATION"
    ],
    "expected_invariant": "D20",
    "result": "CONTAINED"
   },
   "refuted_without_refutation": {
    "category": "false_discovery",
    "defense": "REFUTED needs refuting evidence",
    "evidence": [
     "REFUTING_EVIDENCE_REQUIRED"
    ],
    "expected_invariant": "D11",
    "result": "CONTAINED"
   },
   "reviewer_fabricates_simulation": {
    "category": "fabricated_experiment",
    "defense": "reviewers cannot produce simulations",
    "evidence": [
     "NO_RESEARCH_AUTHORITY:simulate"
    ],
    "expected_invariant": "D58",
    "result": "CONTAINED"
   },
   "revoked_producer_credential": {
    "category": "fabricated_experiment",
    "defense": "revoked credential cannot produce evidence",
    "evidence": [
     "EVIDENCE_PRODUCER_REVOKED",
     "NO_RESEARCH_AUTHORITY:experiment"
    ],
    "expected_invariant": "D14",
    "result": "CONTAINED"
   },
   "simulated_intervention_as_observed": {
    "category": "simulation_laundering",
    "defense": "observed intervention needs a controlled experiment",
    "evidence": [
     "OBSERVED_INTERVENTION_REQUIRES_CONTROLLED_EXPERIMENT"
    ],
    "expected_invariant": "D16",
    "result": "CONTAINED"
   },
   "simulated_without_simulation": {
    "category": "fabricated_experiment",
    "defense": "SIMULATED needs simulation evidence",
    "evidence": [
     "SIMULATION_EVIDENCE_REQUIRED"
    ],
    "expected_invariant": "D54",
    "result": "CONTAINED"
   },
   "simulation_experiment_in_lab": {
    "category": "simulation_laundering",
    "defense": "simulation experiments sandboxed",
    "evidence": [
     "SIMULATION_OUTSIDE_SANDBOX"
    ],
    "expected_invariant": "D16",
    "result": "CONTAINED"
   },
   "simulation_labelled_observational": {
    "category": "simulation_laundering",
    "defense": "simulation data labelled simulated",
    "evidence": [
     "SIMULATION_MUST_BE_LABELLED_SIMULATED"
    ],
    "expected_invariant": "D16",
    "result": "CONTAINED"
   },
   "simulation_only_support": {
    "category": "simulation_laundering",
    "defense": "simulation alone never supports",
    "evidence": {
     "authority": "NONE",
     "confidence": 0.6875,
     "hypothesis": "h1",
     "independent_groups": 6,
     "independent_replications": 0,
     "observational_contradiction": 0,
     "observational_support": 0,
     "quality": "LOW_CONFIDENCE",
     "simulation_only": true,
     "third_party_reproductions": 0,
     "uncertainty": 0.625
    },
    "expected_invariant": "D5",
    "result": "CONTAINED"
   },
   "simulation_relabelled_as_observation": {
    "category": "simulation_laundering",
    "defense": "method class bound into signature",
    "evidence": [
     "EVIDENCE_SIGNATURE_INVALID",
     "NO_RESEARCH_AUTHORITY:experiment"
    ],
    "expected_invariant": "D16",
    "result": "CONTAINED"
   },
   "simulation_run_in_field": {
    "category": "simulation_laundering",
    "defense": "simulations only in sandboxes",
    "evidence": [
     "SIMULATION_OUTSIDE_SANDBOX"
    ],
    "expected_invariant": "D92",
    "result": "CONTAINED"
   },
   "terminated_experiment_results_used": {
    "category": "fabricated_experiment",
    "defense": "terminated runs yield no observations; the failure is kept",
    "evidence": [
     "FORBIDDEN_OPERATION:exfiltrate"
    ],
    "expected_invariant": "D13",
    "result": "CONTAINED"
   },
   "unfalsifiable_hypothesis": {
    "category": "false_discovery",
    "defense": "a hypothesis must predict",
    "evidence": [
     "HYPOTHESIS_UNFALSIFIABLE_NO_PREDICTION"
    ],
    "expected_invariant": "D11",
    "result": "CONTAINED"
   },
   "unregistered_producer": {
    "category": "fabricated_experiment",
    "defense": "producer must be registered",
    "evidence": [
     "EVIDENCE_PRODUCER_UNKNOWN",
     "NO_RESEARCH_AUTHORITY:experiment"
    ],
    "expected_invariant": "D53",
    "result": "CONTAINED"
   }
  },
  "total": 32
 },
 "demonstration": {
  "all_self_authorization_refused": true,
  "all_steps_ok": true,
  "authority": "NONE",
  "execution": "E21 -> E20 -> E19 -> E8 (in-process reference kernel)",
  "note": "a deterministic synthetic problem; it demonstrates governance of the research process, not a scientific result",
  "problem": "synthetic routing-rule comparison",
  "schema": "cain42.e21.research-demonstration.v1",
  "self_authorization_attempts": [
   {
    "attempt": "simulation_as_observation",
    "reasons": [
     "DISCOVERY_TRANSITION_REFUSED:PROPOSED->OBSERVED",
     "OBSERVATIONAL_EVIDENCE_REQUIRED"
    ],
    "refused": true
   },
   {
    "attempt": "experiment_modifies_policy",
    "reasons": [
     "FORBIDDEN_OPERATION:modify_policy"
    ],
    "refused": true
   },
   {
    "attempt": "false_replication",
    "reasons": [
     "NO_RESEARCH_AUTHORITY:replicate",
     "REPLICATION_NOT_INDEPENDENT"
    ],
    "refused": true
   },
   {
    "attempt": "premature_collapse_of_alternatives",
    "reasons": [
     "NO_RESEARCH_AUTHORITY:hypothesize",
     "PREMATURE_COLLAPSE_REFUSED"
    ],
    "refused": true
   },
   {
    "attempt": "support_without_adversarial_challenge",
    "reasons": [
     "ADVERSARIAL_CHALLENGE_REQUIRED",
     "PEER_REVIEW_REQUIRED"
    ],
    "refused": true
   },
   {
    "attempt": "researcher_reviews_own_work",
    "reasons": [
     "NO_RESEARCH_AUTHORITY:review",
     "REVIEWER_IS_A_RESEARCHER",
     "REVIEWER_SHARES_CONTROLLER_WITH_RESEARCHER"
    ],
    "refused": true
   },
   {
    "attempt": "self_confirmation",
    "reasons": [
     "NO_RESEARCH_AUTHORITY:review",
     "RESEARCHER_CANNOT_CONFIRM_OWN_DISCOVERY"
    ],
    "refused": true
   },
   {
    "attempt": "self_asserted_confidence",
    "reasons": [
     "CONFIDENCE_SELF_ASSERTED"
    ],
    "refused": true
   },
   {
    "attempt": "execute_before_promotion",
    "reasons": [
     "CAPABILITY_NOT_PROMOTED"
    ],
    "refused": true
   },
   {
    "attempt": "proposer_tests_own_capability",
    "reasons": [
     "ADVERSARIAL_TESTER_NOT_INDEPENDENT",
     "NO_RESEARCH_AUTHORITY:refute"
    ],
    "refused": true
   },
   {
    "attempt": "promotion_without_quorum",
    "reasons": [
     "CAPABILITY_PROMOTION_REQUIRES_GOVERNANCE"
    ],
    "refused": true
   },
   {
    "attempt": "researchers_sign_their_own_promotion",
    "reasons": [
     "GOVERNANCE_QUORUM_NOT_MET"
    ],
    "refused": true
   },
   {
    "attempt": "outside_limited_deployment",
    "reasons": [
     "LIMITED_DEPLOYMENT_TARGET_EXCEEDED"
    ],
    "refused": true
   },
   {
    "attempt": "scope_escalation_to_dispatch",
    "reasons": [
     "CAPABILITY_SCOPE_EXCEEDED"
    ],
    "refused": true
   },
   {
    "attempt": "operator_outside_limited_deployment",
    "reasons": [
     "LIMITED_DEPLOYMENT_OPERATOR_EXCEEDED"
    ],
    "refused": true
   },
   {
    "attempt": "discovery_as_authority_basis",
    "reasons": [
     "NO_INSTITUTIONAL_AUTHORITY_FOR_CAPABILITY"
    ],
    "refused": true
   }
  ],
  "steps": [
   {
    "detail": {
     "live": [
      "h1",
      "h2",
      "h3"
     ]
    },
    "ok": true,
    "step": "FORMULATE_HYPOTHESES"
   },
   {
    "detail": {
     "tracks": [
      "T-exp",
      "T-rep",
      "T-sim"
     ]
    },
    "ok": true,
    "step": "COMPETING_TRACKS"
   },
   {
    "detail": {
     "controlled": [],
     "sim": []
    },
    "ok": true,
    "step": "DESIGN_EXPERIMENTS"
   },
   {
    "detail": {
     "delay_reduction": 0.1996,
     "execution": "SANDBOXED",
     "label": "SIMULATION"
    },
    "ok": true,
    "step": "RUN_SIMULATIONS"
   },
   {
    "detail": {
     "delay_reduction": 0.1993,
     "label": "CONTROLLED_EXPERIMENT"
    },
    "ok": true,
    "step": "PRODUCE_OBSERVATIONS"
   },
   {
    "detail": {
     "authority": "NONE",
     "confidence": 0.294118,
     "hypothesis": "h2",
     "independent_groups": 2,
     "independent_replications": 0,
     "observational_contradiction": 1,
     "observational_support": 0,
     "quality": "REFUTED",
     "simulation_only": false,
     "third_party_reproductions": 0,
     "uncertainty": 0.588235
    },
    "ok": true,
    "step": "IDENTIFY_CONTRADICTIONS"
   },
   {
    "detail": {
     "competition": {
      "authority": "NONE",
      "blind_replication_broken": [],
      "committed_after_reveal": [],
      "contaminated": [],
      "independent_convergence": [
       [
        "T-exp",
        "T-rep"
       ]
      ],
      "independent_tracks": [
       "T-exp",
       "T-rep"
      ],
      "premature_consensus_prevented": true,
      "problem": "q1",
      "tracks": [
       "T-exp",
       "T-rep",
       "T-sim"
      ]
     },
     "h1": {
      "authority": "NONE",
      "confidence": 0.8,
      "hypothesis": "h1",
      "independent_groups": 3,
      "independent_replications": 1,
      "observational_contradiction": 0,
      "observational_support": 2,
      "quality": "SUPPORTED",
      "simulation_only": false,
      "third_party_reproductions": 0,
      "uncertainty": 0.4
     }
    },
    "ok": true,
    "step": "REPLICATE_RESULT"
   },
   {
    "detail": {
     "kept": [
      "h1"
     ],
     "retired": {
      "h2": "refuted",
      "h3": "refuted"
     }
    },
    "ok": true,
    "step": "REJECT_HYPOTHESIS"
   },
   {
    "detail": {
     "discovery": "SUPPORTED",
     "knowledge": {
      "authority": "NONE",
      "claim": "rule A reduces delay",
      "claim_id": "k1",
      "confidence": 0.8,
      "contradictions": [],
      "dependencies": [],
      "evidence": [
       "e-ctl",
       "e-rep"
      ],
      "method": "controlled comparison + independent replication",
      "quality": "SUPPORTED",
      "source": "d1",
      "state": "SUPPORTED",
      "timestamp": 1001.0,
      "uncertainty": 0.4,
      "version": 1
     }
    },
    "ok": true,
    "step": "CANDIDATE_CAPABILITY"
   },
   {
    "detail": {
     "stage": "QUARANTINED"
    },
    "ok": true,
    "step": "QUARANTINE_CAPABILITY"
   },
   {
    "detail": {
     "stage": "SECURITY_REVIEWED"
    },
    "ok": true,
    "step": "ADVERSARIAL_TEST"
   },
   {
    "detail": {
     "promotion": {
      "grants_authority": false,
      "scope": [
       "route"
      ],
      "targets": [
       "depot-sandbox-1"
      ]
     }
    },
    "ok": true,
    "step": "PROMOTE_AFTER_GOVERNANCE"
   },
   {
    "detail": {
     "commit": "E8",
     "decision": "AUTHORIZED",
     "reasons": []
    },
    "ok": true,
    "step": "BOUNDED_DEMONSTRATION"
   },
   {
    "detail": {
     "abs_error": 0.0002,
     "observed": 0.2002,
     "predicted": 0.2,
     "tolerance": 0.05
    },
    "ok": true,
    "step": "PREDICTION_VS_OUTCOME"
   },
   {
    "detail": {
     "authority": "NONE",
     "claim": "rule A reduces delay",
     "claim_id": "k1",
     "confidence": 0.833333,
     "contradictions": [],
     "dependencies": [],
     "evidence": [
      "e-ctl",
      "e-outcome",
      "e-rep"
     ],
     "method": "controlled comparison + independent replication",
     "quality": "SUPPORTED",
     "source": "d1",
     "state": "SUPPORTED",
     "timestamp": 1006.0,
     "uncertainty": 0.333333,
     "version": 2
    },
    "ok": true,
    "step": "UPDATE_KNOWLEDGE"
   },
   {
    "detail": {
     "failures": [
      {
       "experiment": "x-bad",
       "reasons": [
        "FORBIDDEN_OPERATION:modify_policy"
       ],
       "status": "TERMINATED"
      },
      {
       "hypothesis": "h2",
       "reason": "refuted"
      },
      {
       "hypothesis": "h3",
       "reason": "refuted"
      }
     ],
     "memory_entries": 33
    },
    "ok": true,
    "step": "PRESERVE_FAILED_PATHS"
   },
   {
    "detail": {
     "not_refused": [],
     "refused": [
      "simulation_as_observation",
      "experiment_modifies_policy",
      "false_replication",
      "premature_collapse_of_alternatives",
      "support_without_adversarial_challenge",
      "researcher_reviews_own_work",
      "self_confirmation",
      "self_asserted_confidence",
      "execute_before_promotion",
      "proposer_tests_own_capability",
      "promotion_without_quorum",
      "researchers_sign_their_own_promotion",
      "outside_limited_deployment",
      "scope_escalation_to_dispatch",
      "operator_outside_limited_deployment",
      "discovery_as_authority_basis"
     ]
    },
    "ok": true,
    "step": "NO_SELF_AUTHORIZATION"
   }
  ]
 },
 "invariants": [
  {
   "detail": [
    [
     "CAPABILITY_NOT_PROMOTED"
    ],
    [
     "NO_INSTITUTIONAL_AUTHORITY_FOR_CAPABILITY"
    ]
   ],
   "holds": true,
   "id": "D1",
   "invariant": "discovery cannot create authority"
  },
  {
   "detail": [
    [
     "EVIDENCE_SIGNATURE_INVALID",
     "NO_RESEARCH_AUTHORITY:experiment"
    ],
    [
     "SIMULATION_MUST_BE_LABELLED_SIMULATED"
    ],
    [
     "STAGE_MASQUERADE:SIMULATION->OBSERVATION"
    ]
   ],
   "holds": true,
   "id": "D16",
   "invariant": "simulation remains simulation"
  },
  {
   "detail": [
    [
     "DISCOVERY_TRANSITION_REFUSED:PROPOSED->VERIFIED",
     "THIRD_PARTY_REPRODUCTION_REQUIRED"
    ],
    [
     "KNOWLEDGE_EVIDENCE_REQUIRED"
    ]
   ],
   "holds": true,
   "id": "D54",
   "invariant": "missing evidence does not become verified"
  },
  {
   "detail": [
    [
     "ADVERSARIAL_CHALLENGE_REQUIRED",
     "DISCOVERY_TRANSITION_REFUSED:PROPOSED->SUPPORTED",
     "EVIDENCE_DOES_NOT_SUPPORT:UNKNOWN",
     "INDEPENDENT_REPLICATION_REQUIRED",
     "PEER_REVIEW_REQUIRED"
    ],
    [
     "OBSERVATIONAL_EVIDENCE_REQUIRED"
    ]
   ],
   "holds": true,
   "id": "D123",
   "invariant": "the discovery lifecycle has no shortcuts"
  }
 ],
 "schema": "cain42.e21.discovery-results.v1"
}