{
  "schema_version": "amoeba.paper_evidence_ledger.v1",
  "paper_id": "amoeba_empirical_v0_1",
  "title": "Equal Evidence, Different Outcomes: A Process-Isolated Evaluation of Governed External Reasoning Around Small Language Models",
  "last_updated": "2026-07-14",
  "repository": "https://github.com/headsoftsoftware/amoeba_core_v2.git",
  "repository_root": "H:/amoeba_core_v2",
  "evidence_cutoff_commit": "d85ccccaacfd733e8761cff614f3368226d4f890",
  "official_score_claims_enabled": false,
  "independence_policy": "Replay phases, derived policy analyses, ablations, and post-hoc verifiers are never counted as new independent target samples.",
  "entries": [
    {
      "id": "E1",
      "title": "Process-isolated live equal-evidence causal target",
      "paper_role": "primary_headline_result",
      "evidence_class": "live_process_isolated_causal_mechanism",
      "independence_group": "isolated_live_equal_evidence_target_20260714",
      "is_independent_sample": true,
      "source_docs": [
        "docs/ISOLATED_LIVE_EQUAL_EVIDENCE_PREREGISTRATION_2026-07-14.md",
        "docs/ISOLATED_LIVE_EQUAL_EVIDENCE_TARGET_RESULT_2026-07-14.md",
        "docs/ISOLATED_LIVE_EQUAL_EVIDENCE_VERIFICATION_RESULT_2026-07-14.md"
      ],
      "artifact_commits": {
        "implementation": "5e72cc1",
        "frozen_receipt": "ee75ff4",
        "result": "8426a6f",
        "independent_verifier": "03b7b11",
        "verification_result": "0248721"
      },
      "artifacts": [
        {
          "path": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "sha256": "2050796f3bf16a72b97fa30b337b472c41f135c9bcc95050b5d5da4944e8b3ef"
        },
        {
          "path": "reports/isolated_live_equal_evidence_target_verification_20260714.json",
          "sha256": "7887e3335bfa9b8fec3c2f1df425c7537407b2a5e51fea3c5a3533597f2f7b56"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "case_count",
          "value": 48
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "phase_count",
          "value": 2
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "result_row_count",
          "value": 288
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "phase_summaries.phase_a.raw_proposal.correct",
          "value": 18
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "phase_summaries.phase_a.flat_equal_evidence.correct",
          "value": 14
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "phase_summaries.phase_a.governed_runtime.correct",
          "value": 48
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "phase_summaries.phase_b.raw_proposal.correct",
          "value": 18
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "phase_summaries.phase_b.flat_equal_evidence.correct",
          "value": 14
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "phase_summaries.phase_b.governed_runtime.correct",
          "value": 48
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "paired_comparisons.phase_a.wins",
          "value": 34
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "paired_comparisons.phase_a.losses",
          "value": 0
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "paired_comparisons.phase_a.ties",
          "value": 14
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "paired_comparisons.phase_a.exact_two_sided_sign_p.decimal",
          "value": "1.16415321826934814e-10"
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "raw_governed_proposal_agreement.phase_a.matched",
          "value": 48
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "flat_governed_evidence_and_work_match.phase_a.exact",
          "value": true
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "phase_summaries.phase_a.flat_equal_evidence.input_tokens",
          "value": 32477
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "phase_summaries.phase_a.governed_runtime.input_tokens",
          "value": 28312
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "replay_comparison.exact",
          "value": true
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "replay_comparison.compared",
          "value": 144
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "isolation_proof.passed",
          "value": true
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "isolation_proof.required_epoch_count",
          "value": 6
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "target_raw_model_text_persisted",
          "value": false
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_result_20260714.json",
          "path": "authorized_official_evaluator_reads",
          "value": 0
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_verification_20260714.json",
          "path": "all_checks_passed",
          "value": true
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_verification_20260714.json",
          "path": "target_reexecuted",
          "value": false
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_verification_20260714.json",
          "path": "campaign_generator_called",
          "value": false
        },
        {
          "artifact": "reports/isolated_live_equal_evidence_target_verification_20260714.json",
          "path": "model_or_provider_called",
          "value": false
        }
      ],
      "supported_claim": "On one preregistered generated 48-case family, the governed deterministic route changed outcomes after identical Gemma 1B proposals and outperformed a flat same-model route given exactly the same evidence and matched work under fresh-process isolation.",
      "excluded_claims": [
        "general superiority over RAG",
        "external benchmark superiority",
        "learned transfer",
        "frontier-model superiority",
        "broad autonomous scientific reasoning",
        "a second independent n=48 replication from the replay phase"
      ],
      "official_score_claim": false
    },
    {
      "id": "E2",
      "title": "Full local HumanEval primary-route capability result",
      "paper_role": "secondary_capability_demonstration",
      "evidence_class": "recognized_dataset_local_nonofficial_capability",
      "independence_group": "humaneval_primary_route_20260709",
      "is_independent_sample": true,
      "source_docs": [
        "docs/HUMANEVAL_PRIMARY_ROUTE_PASS_AT_ONE_2026-07-09.md"
      ],
      "artifact_commits": {
        "report": "3a29b6fce86943488c6badd5d9c54b4486bbd1be"
      },
      "artifacts": [
        {
          "path": "reports/humaneval_primary_route_full_summary_20260709.json",
          "sha256": "e68ff737ec1578d8f91d054f5d2ecd2651a7b10f978b02edfa38cc6cc94750fe"
        },
        {
          "path": "reports/humaneval_official_style_claim_primary_route_full_20260709.json",
          "sha256": "dcc235e45bbb72797a4aad2eba9ee4a6b060503495fee923ee2910098484d352"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "problem_count",
          "value": 164
        },
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "raw_pass_count",
          "value": 129
        },
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "amoeba_pass_count",
          "value": 164
        },
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "pass_rate_delta",
          "operator": "approx_equals",
          "value": 0.21341463414634146,
          "tolerance": 1e-15
        },
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "positive_delta_count",
          "value": 35
        },
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "negative_delta_count",
          "value": 0
        },
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "candidate_selection",
          "value": null
        },
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "knowledge_access_lane",
          "value": "closed_book"
        },
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "official_score_claim",
          "value": false
        },
        {
          "artifact": "reports/humaneval_primary_route_full_summary_20260709.json",
          "path": "official_run_readiness.status",
          "value": "blocked"
        },
        {
          "artifact": "reports/humaneval_official_style_claim_primary_route_full_20260709.json",
          "path": "summary.same_circumstances.selection_or_retry_used_for_pass_at_1",
          "value": false
        },
        {
          "artifact": "reports/humaneval_official_style_claim_primary_route_full_20260709.json",
          "path": "summary.claim.official_leaderboard_score_claim_enabled",
          "value": false
        },
        {
          "artifact": "reports/humaneval_official_style_claim_primary_route_full_20260709.json",
          "path": "summary.dataset_ref.sha256",
          "value": "b796127e635a67f93fb35c04f4cb03cf06f38c8072ee7cee8833d7bee06979ef"
        }
      ],
      "supported_claim": "On the local closed-book 164-task HumanEval artifact, one raw completion per task passed 129/164 and one Amoeba primary-route completion per task passed 164/164 without post-test candidate selection.",
      "excluded_claims": [
        "official HumanEval leaderboard score",
        "official sandbox parity",
        "untouched benchmark generalization",
        "frontier-model comparison",
        "proof that every intervention was learned rather than benchmark-developed"
      ],
      "official_score_claim": false
    },
    {
      "id": "E3",
      "title": "Synthetic held-out method-transfer and retrieval comparator",
      "paper_role": "retrieval_comparator_and_limit",
      "evidence_class": "synthetic_heldout_method_transfer",
      "independence_group": "humaneval_shaped_transfer_72_20260705",
      "is_independent_sample": true,
      "source_docs": [
        "docs/HUMANEVAL_LAYER_HYPOTHESES.md"
      ],
      "artifacts": [
        {
          "path": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "sha256": "243d1e3aab1961af07bb8fdbee9ce284ff33d4fa4941b7966363d3b22092f703"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "path": "summary.large_heldout.heldout_case_count",
          "value": 72
        },
        {
          "artifact": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "path": "summary.large_heldout.raw_pass_count",
          "value": 10
        },
        {
          "artifact": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "path": "summary.large_heldout.flat_same_evidence_pass_count",
          "value": 39
        },
        {
          "artifact": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "path": "summary.large_heldout.surface_retrieval_top2_pass_count",
          "value": 69
        },
        {
          "artifact": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "path": "summary.large_heldout.surface_retrieval_top4_pass_count",
          "value": 72
        },
        {
          "artifact": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "path": "summary.large_heldout.amoeba_typed_substrate_pass_count",
          "value": 72
        },
        {
          "artifact": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "path": "summary.large_heldout.proof_gate_ablation_pass_count",
          "value": 0
        },
        {
          "artifact": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "path": "summary.large_heldout.selector_ablation_pass_count",
          "value": 1
        },
        {
          "artifact": "reports/humaneval_capability_layer_scorecard_2026-07-05.json",
          "path": "summary.official_score_claim",
          "value": false
        }
      ],
      "supported_claim": "On one synthetic 72-case held-out aggregate, method evidence, retrieval, and typed selection were measurable, but a top-4 surface retriever tied typed Amoeba at 72/72.",
      "excluded_claims": [
        "broad superiority over RAG",
        "recognized benchmark transfer",
        "independently sourced holdout",
        "official HumanEval score"
      ],
      "official_score_claim": false
    },
    {
      "id": "E4",
      "title": "Untouched whole-family HumanEval-shaped transfer replay",
      "paper_role": "negative_or_limited_transfer_result",
      "evidence_class": "frozen_project_untouched_family_holdout",
      "independence_group": "humaneval_untouched_transfer_v2_20260714",
      "is_independent_sample": true,
      "source_docs": [
        "docs/HUMANEVAL_UNTOUCHED_TRANSFER_RESULT_2026-07-14.md"
      ],
      "artifact_commits": {
        "preregistration": "e8bf8a6f320b44bd5f7ba2f71ee24267824ae19b",
        "implementation": "160747e3574d6c6a075f3d9d9eebdc3d675eb156"
      },
      "artifacts": [
        {
          "path": "reports/benchmark_campaigns/human_eval__untouched_transfer__target_evaluation_v2_20260714.json",
          "sha256": "64fec28ce7adaa1afe6b99f0708cdc98f827557b8f0123dd6fee0d4a18deabf7"
        },
        {
          "path": "reports/benchmark_campaigns/human_eval__empirical_policy_holdout__qwen2-5-coder-14b__n16__untouched-transfer-v2-20260714__report.json",
          "sha256": "801985dfe760a22d0bfc340c3c61436c3df105788a70b5b3b377acc509172d41"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/benchmark_campaigns/human_eval__untouched_transfer__target_evaluation_v2_20260714.json",
          "path": "status",
          "value": "completed"
        },
        {
          "artifact": "reports/benchmark_campaigns/human_eval__untouched_transfer__target_evaluation_v2_20260714.json",
          "path": "target_campaign.task_count",
          "value": 16
        },
        {
          "artifact": "reports/benchmark_campaigns/human_eval__untouched_transfer__target_evaluation_v2_20260714.json",
          "path": "target_campaign.matrix.complete",
          "value": true
        },
        {
          "artifact": "reports/benchmark_campaigns/human_eval__untouched_transfer__target_evaluation_v2_20260714.json",
          "path": "policy_conditions.global_source_best.success_count",
          "value": 14
        },
        {
          "artifact": "reports/benchmark_campaigns/human_eval__untouched_transfer__target_evaluation_v2_20260714.json",
          "path": "policy_conditions.hand_static.success_count",
          "value": 12
        },
        {
          "artifact": "reports/benchmark_campaigns/human_eval__untouched_transfer__target_evaluation_v2_20260714.json",
          "path": "primary_comparisons.hand_static.exact_two_sided_sign_test_p",
          "value": 0.5
        },
        {
          "artifact": "reports/benchmark_campaigns/human_eval__untouched_transfer__target_evaluation_v2_20260714.json",
          "path": "target_outcomes_used_for_refit",
          "value": false
        },
        {
          "artifact": "reports/benchmark_campaigns/human_eval__untouched_transfer__target_evaluation_v2_20260714.json",
          "path": "task_or_family_identifiers_visible_to_policy",
          "value": false
        },
        {
          "artifact": "reports/benchmark_campaigns/human_eval__empirical_policy_holdout__qwen2-5-coder-14b__n16__untouched-transfer-v2-20260714__report.json",
          "path": "result_error_count",
          "value": 0
        },
        {
          "artifact": "reports/benchmark_campaigns/human_eval__empirical_policy_holdout__qwen2-5-coder-14b__n16__untouched-transfer-v2-20260714__report.json",
          "path": "official_score.enabled",
          "value": false
        }
      ],
      "supported_claim": "A frozen selected policy reached 14/16 versus 12/16 for the static policy on one project-untouched family-disjoint holdout, but the paired difference was not statistically significant (p=0.5).",
      "excluded_claims": [
        "globally pristine HumanEval transfer",
        "significant selected-policy superiority",
        "generic pretraining transfer",
        "official pass@1",
        "independent evidence that evolution improved primary-route accuracy"
      ],
      "official_score_claim": false
    },
    {
      "id": "E5",
      "title": "Knowledge-independent deduction component battery",
      "paper_role": "mechanism_ablation",
      "evidence_class": "generated_answer_hidden_symbolic_ablation",
      "independence_group": "knowledge_independent_deduction_20260714",
      "is_independent_sample": true,
      "source_docs": [
        "docs/KNOWLEDGE_INDEPENDENT_DEDUCTION_BATTERY_RESULT_2026-07-14.md"
      ],
      "artifact_commits": {
        "implementation": "7b4fa58ba826e1cba739159c714ec9eef3be09a8",
        "holdout_freeze": "6008a210b0b6cea46ed661e234f0cf2439ad489c"
      },
      "artifacts": [
        {
          "path": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "sha256": "be9a63d9af1cf78ab85f05ef4c4805ab209de475877380bab718dae9b7a2f4eb"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "condition_results.static.correct",
          "value": 2
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "condition_results.proof_only.correct",
          "value": 10
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "condition_results.flat_retry.correct",
          "value": 20
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "condition_results.gap_obligation.correct",
          "value": 24
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "condition_results.hypothesis_frontier.correct",
          "value": 36
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "condition_results.verifier_frontier.correct",
          "value": 60
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "condition_results.full_localized.correct",
          "value": 72
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "paired_comparisons.full_localized_vs_verifier_frontier.paired_wins",
          "value": 12
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "paired_comparisons.full_localized_vs_verifier_frontier.paired_losses",
          "value": 0
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "paired_comparisons.full_localized_vs_verifier_frontier.exact_two_sided_sign_p.decimal",
          "value": "4.882812500000000000E-4"
        },
        {
          "artifact": "reports/knowledge_independent_deduction_analysis_20260714.json",
          "path": "all_holdout_gates_passed",
          "value": true
        }
      ],
      "supported_claim": "On a designed 72-case answer-hidden symbolic battery, obligation, hypothesis, verifier, and localization components made separable contributions; the full route reached 72/72 versus 60/72 for the strongest ablation.",
      "excluded_claims": [
        "live-model reasoning",
        "external benchmark validity",
        "broad scientific intelligence",
        "real-world task-family generalization"
      ],
      "official_score_claim": false
    },
    {
      "id": "E6",
      "title": "Context-scoped marginal route admission",
      "paper_role": "derived_governance_policy",
      "evidence_class": "post_hoc_derived_policy_analysis",
      "independence_group": "isolated_live_equal_evidence_target_20260714",
      "is_independent_sample": false,
      "derived_from": [
        "E1"
      ],
      "source_docs": [
        "docs/MARGINAL_ROUTE_ADMISSION_RESULT_2026-07-14.md"
      ],
      "artifact_commits": {
        "implementation": "7ba61cb7b4b10f777d0b259771543d00fcc015b1",
        "result": "d85cccc"
      },
      "artifacts": [
        {
          "path": "reports/route_value_admission_target_result_20260714.json",
          "sha256": "141940a6a79a183b0594ecb713f1dc84fe3b33fb6b18b764b78b488350d4feb5"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/route_value_admission_target_result_20260714.json",
          "path": "paired_case_count",
          "value": 96
        },
        {
          "artifact": "reports/route_value_admission_target_result_20260714.json",
          "path": "decisions.governed_runtime.decision",
          "value": "admit"
        },
        {
          "artifact": "reports/route_value_admission_target_result_20260714.json",
          "path": "decisions.governed_runtime.paired.wins",
          "value": 60
        },
        {
          "artifact": "reports/route_value_admission_target_result_20260714.json",
          "path": "decisions.governed_runtime.paired.losses",
          "value": 0
        },
        {
          "artifact": "reports/route_value_admission_target_result_20260714.json",
          "path": "decisions.flat_equal_evidence.decision",
          "value": "preserve_baseline"
        },
        {
          "artifact": "reports/route_value_admission_target_result_20260714.json",
          "path": "decisions.flat_equal_evidence.paired.wins",
          "value": 10
        },
        {
          "artifact": "reports/route_value_admission_target_result_20260714.json",
          "path": "decisions.flat_equal_evidence.paired.losses",
          "value": 18
        },
        {
          "artifact": "reports/route_value_admission_target_result_20260714.json",
          "path": "new_model_calls",
          "value": 0
        },
        {
          "artifact": "reports/route_value_admission_target_result_20260714.json",
          "path": "new_target_executions",
          "value": 0
        }
      ],
      "supported_claim": "The frozen E1 history can be converted into a context-keyed admission decision that admits the governed route and preserves raw ahead of the inconclusive flat route.",
      "excluded_claims": [
        "new experimental sample",
        "proof that flat evidence is harmful",
        "global route superiority",
        "authorization to rerun or relabel E1"
      ],
      "official_score_claim": false
    },
    {
      "id": "E7",
      "title": "Noisy latent mechanism and governance controls",
      "paper_role": "negative_accuracy_result_and_governance_mechanism",
      "evidence_class": "generated_noisy_latent_ablation",
      "independence_group": "noisy_latent_mechanism_20260714",
      "is_independent_sample": true,
      "source_docs": [
        "docs/NOISY_LATENT_MECHANISM_RESULT_2026-07-14.md"
      ],
      "artifacts": [
        {
          "path": "reports/noisy_latent_mechanism_result_20260714.json",
          "sha256": "ca190b24b5cfba58a45a73f04bbece83ea0663406a2866294277985be17376b7"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "case_count",
          "value": 192
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "result_row_count",
          "value": 960
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "summary.conditions.static_prior.clean_task_correct",
          "value": 24
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "summary.conditions.fixed_order_bayes.clean_task_correct",
          "value": 35
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "summary.conditions.plain_same_evidence_map.clean_task_correct",
          "value": 86
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "summary.conditions.adaptive_ungoverned.clean_task_correct",
          "value": 85
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "summary.conditions.governed_adaptive.clean_task_correct",
          "value": 86
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "summary.conditions.adaptive_ungoverned.special_governance_passed",
          "value": 16
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "summary.conditions.governed_adaptive.special_governance_passed",
          "value": 96
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "deterministic_replay_match",
          "value": true
        },
        {
          "artifact": "reports/noisy_latent_mechanism_result_20260714.json",
          "path": "model_calls",
          "value": 0
        }
      ],
      "supported_claim": "On one generated noisy-latent battery, governed and plain exact MAP tied at 86/96 clean cases, while governance passed all 96 special authority controls versus 16/96 for the ungated adaptive route.",
      "excluded_claims": [
        "accuracy gain over an exact same-evidence calculator",
        "live-model uncertainty reasoning",
        "external benchmark generalization"
      ],
      "official_score_claim": false
    },
    {
      "id": "E8",
      "title": "Evolutionary typed-policy compiler mechanism",
      "paper_role": "exploratory_mechanism_appendix",
      "evidence_class": "generated_policy_search_ablation",
      "independence_group": "evolutionary_policy_compiler_20260714",
      "is_independent_sample": true,
      "source_docs": [
        "docs/EVOLUTIONARY_POLICY_COMPILER_RESULT_2026-07-14.md"
      ],
      "artifact_commits": {
        "preregistered_implementation": "3561294e5bb27c3aed635d11954aabf7936cfaa7"
      },
      "artifacts": [
        {
          "path": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "sha256": "76af9896f2eb62431f3830fbfbd1983c11ab3a84390432ee259ee991da38ebd5"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "path": "mechanism_endpoint.full_holdout_success_rate",
          "value": 1.0
        },
        {
          "artifact": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "path": "mechanism_endpoint.baseline_holdout_success_rate",
          "value": 0.25
        },
        {
          "artifact": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "path": "mechanism_endpoint.global_deep_holdout_success_rate",
          "value": 0.5
        },
        {
          "artifact": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "path": "evolution_endpoint.treatment_full_holdout_count",
          "value": 12
        },
        {
          "artifact": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "path": "evolution_endpoint.comparator_full_holdout_count",
          "value": 2
        },
        {
          "artifact": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "path": "evolution_endpoint.paired_comparison.better",
          "value": 10
        },
        {
          "artifact": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "path": "evolution_endpoint.paired_comparison.worse",
          "value": 0
        },
        {
          "artifact": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "path": "evolution_endpoint.paired_comparison.exact_two_sided_sign_p",
          "value": 0.001953125
        },
        {
          "artifact": "reports/evolutionary_policy_compiler_measured_summary_20260714.json",
          "path": "model_calls",
          "value": 0
        }
      ],
      "supported_claim": "On a favorable generated policy-search landscape, archive-guided crossover plus mutation found the full held-out policy in 12/12 repeats versus 2/12 for the strongest equal-budget comparator.",
      "excluded_claims": [
        "recognized benchmark gain",
        "live model routing gain",
        "general evolutionary superiority",
        "autonomous promotion",
        "independent task-family replication"
      ],
      "official_score_claim": false
    },
    {
      "id": "E9",
      "title": "Live bounded scientific-method gate session",
      "paper_role": "secondary_live_mechanism",
      "evidence_class": "live_process_isolated_bounded_session",
      "independence_group": "scientific_method_live_session_20260702",
      "is_independent_sample": true,
      "source_docs": [
        "docs/CRITICAL_EVIDENCE_INDEX.md"
      ],
      "artifacts": [
        {
          "path": "reports/scientific_method_model_session_qwen25_coder14b_isolated_provider_after_backup_2026-07-02.json",
          "sha256": "d31d3ae4c7d661ac8e1c63889a36a97b686abd13e2a4659ea36e3aaf0b37bd81"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/scientific_method_model_session_qwen25_coder14b_isolated_provider_after_backup_2026-07-02.json",
          "path": "summary.raw_score",
          "value": 0.3333
        },
        {
          "artifact": "reports/scientific_method_model_session_qwen25_coder14b_isolated_provider_after_backup_2026-07-02.json",
          "path": "summary.full_score",
          "value": 1.0
        },
        {
          "artifact": "reports/scientific_method_model_session_qwen25_coder14b_isolated_provider_after_backup_2026-07-02.json",
          "path": "summary.control_pass_count",
          "value": 0
        },
        {
          "artifact": "reports/scientific_method_model_session_qwen25_coder14b_isolated_provider_after_backup_2026-07-02.json",
          "path": "summary.model_call_count",
          "value": 6
        },
        {
          "artifact": "reports/scientific_method_model_session_qwen25_coder14b_isolated_provider_after_backup_2026-07-02.json",
          "path": "provider_isolation.provider_process_isolation",
          "value": "verified"
        },
        {
          "artifact": "reports/scientific_method_model_session_qwen25_coder14b_isolated_provider_after_backup_2026-07-02.json",
          "path": "provider_isolation.epoch_count",
          "value": 6
        }
      ],
      "supported_claim": "A live Qwen 14B model was constrained by substrate-owned hypothesis, prediction, source, proof, confidence, cost, and selector gates in one bounded process-isolated session.",
      "excluded_claims": [
        "open-ended autonomous science",
        "broad research ability",
        "independent experiment design by the model",
        "recognized benchmark performance"
      ],
      "official_score_claim": false
    },
    {
      "id": "E10",
      "title": "Frozen public-source prestudy: Python 3.14 t-strings",
      "paper_role": "small_n_public_source_replication",
      "evidence_class": "prestudy_frozen_exact_url",
      "independence_group": "prestudy_python_tstrings_20260629",
      "is_independent_sample": true,
      "source_docs": [
        "docs/CRITICAL_EVIDENCE_INDEX.md"
      ],
      "artifacts": [
        {
          "path": "reports/prestudy_frozen_python_tstrings_exact_url_qwen25_coder14b_2026-06-29.json",
          "sha256": "31bc31345249e7947aea1df4a87894e189dbe7a485c6477476cdd4a28b237833"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/prestudy_frozen_python_tstrings_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.question_count",
          "value": 3
        },
        {
          "artifact": "reports/prestudy_frozen_python_tstrings_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.raw_pass_count",
          "value": 0
        },
        {
          "artifact": "reports/prestudy_frozen_python_tstrings_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.amoeba_pass_count",
          "value": 3
        },
        {
          "artifact": "reports/prestudy_frozen_python_tstrings_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.evidence_removed_control_pass_count",
          "value": 0
        },
        {
          "artifact": "reports/prestudy_frozen_python_tstrings_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.wrong_study_control_pass_count",
          "value": 0
        },
        {
          "artifact": "reports/prestudy_frozen_python_tstrings_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "contamination_report.status",
          "value": "pass"
        }
      ],
      "supported_claim": "In one three-question exact-URL prestudy smoke, frozen public-source memory changed same-model outcomes from 0/3 to 3/3 while evidence-removed, wrong-study, and unrelated-study controls remained 0/3.",
      "excluded_claims": [
        "large-sample public-source generalization",
        "recognized benchmark performance",
        "closed-book improvement",
        "broad web autonomy"
      ],
      "official_score_claim": false
    },
    {
      "id": "E11",
      "title": "Frozen public-source prestudy: Kubernetes stop signals",
      "paper_role": "small_n_cross_domain_public_source_replication",
      "evidence_class": "prestudy_frozen_exact_url",
      "independence_group": "prestudy_kubernetes_stop_signals_20260629",
      "is_independent_sample": true,
      "source_docs": [
        "docs/CRITICAL_EVIDENCE_INDEX.md"
      ],
      "artifacts": [
        {
          "path": "reports/prestudy_frozen_kubernetes_stop_signals_exact_url_qwen25_coder14b_2026-06-29.json",
          "sha256": "b464d07413e43a11bc6b580dc4962eea73918f875ffc6e1dd241e45413c57391"
        }
      ],
      "assertions": [
        {
          "artifact": "reports/prestudy_frozen_kubernetes_stop_signals_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.question_count",
          "value": 3
        },
        {
          "artifact": "reports/prestudy_frozen_kubernetes_stop_signals_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.raw_pass_count",
          "value": 0
        },
        {
          "artifact": "reports/prestudy_frozen_kubernetes_stop_signals_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.amoeba_pass_count",
          "value": 3
        },
        {
          "artifact": "reports/prestudy_frozen_kubernetes_stop_signals_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.evidence_removed_control_pass_count",
          "value": 0
        },
        {
          "artifact": "reports/prestudy_frozen_kubernetes_stop_signals_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "report.wrong_study_control_pass_count",
          "value": 0
        },
        {
          "artifact": "reports/prestudy_frozen_kubernetes_stop_signals_exact_url_qwen25_coder14b_2026-06-29.json",
          "path": "contamination_report.status",
          "value": "pass"
        }
      ],
      "supported_claim": "A second three-question exact-URL prestudy smoke in a different source domain reproduced the 0/3 raw to 3/3 frozen-memory pattern with three clean controls.",
      "excluded_claims": [
        "large-sample cross-domain replication",
        "recognized benchmark performance",
        "model portability",
        "broad web autonomy"
      ],
      "official_score_claim": false
    }
  ]
}