File size: 16,591 Bytes
177308a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 | {
"paper_id": "mOcTXKawFY",
"upstream_pin": {
"source": "https://arxiv.org/html/2606.05689v1",
"version": "2606.05689v1",
"sha256": "9b3075a937ce509d5883709a90fba12a35111d6b20f92e033d3e64636728c070"
},
"release_quality_gate": {
"status": "pass_12_of_12_semantic_v4",
"registered_claims": 6,
"supported_by_independent_evidence": 6,
"literal_falsifications": 1,
"expected_verified_points": 12,
"judge_target": "verified_or_literal_falsification",
"formula_only_support_counted": false,
"proxy_support_counted": false,
"algebraic_bound_substitution_counted": false,
"semantic_quality_gate_version": 4,
"direct_rate_claims": 0,
"independent_seeded_trials": 1810,
"exact_derivation_cells": 1822710
},
"claims": [
{
"claim": 1,
"assessment": "verified",
"evidence_tier": "literal_claim_experiment",
"claim_object_match": "exact",
"registered_system_executed": true,
"paper_or_released_scale": true,
"actual_model_or_dataset_used": true,
"destructive_control_executed": true,
"literal_claim": "Definition 1 formalizes an evolutionary selection model as a DAG G^(T) over trait variables X^(0)...X^(T), heritable factors epsilon^(0)...epsilon^(T), and reproduction/selection indicators S^(0)...S^(T-1), distinguishing it from one-shot static selection models (Section 2, Definition 1).",
"paper_native_mechanism": "The executable constructs Definition 1's fully unrolled evolutionary DAG with the printed trait, selection, factor-to-trait, and inheritance edge families.",
"independent_oracle": "Closed-form node and edge counts plus independent topological acyclicity and edge-family inventories are evaluated for every constructed graph.",
"not_proxy_reason": "The audited object is the literal G^(T) graph over all three registered node families; a static DAG or generic layered-network surrogate is not substituted.",
"native_scale_justification": "All 8,688 four-trait static models are unrolled at T=1,2,3 for 26,064 exact constructions, followed by 60 larger constructions through d=20 and T=5.",
"source_locator": "arXiv 2606.05689v1, Section 2, Definition 1 and its four-item edge inventory.",
"independent_evidence": ["outputs/exact_graph_audit.json", "outputs/results.json"],
"executed_outputs": ["outputs/exact_graph_audit.json", "outputs/results.json"],
"oracle_artifacts": ["outputs/oracle_gates.json", "SOURCE_EXCERPTS.md"],
"control_artifacts": ["outputs/destructive_controls.json"],
"destructive_or_boundary_control": "One inheritance edge is deleted from an otherwise valid G^(3); the exact edge-count and edge-family checker rejects the mutant.",
"result": "26,064/26,064 exhaustive constructions and 60/60 larger constructions are acyclic and match the exact node count, edge count, and required edge families.",
"limitation": "The construction audit verifies the formal graph object, not biological adequacy of independent component-wise heritable factors.",
"scope_boundary": "Verification covers Definition 1's graph construction and its distinction from a one-shot model, not arbitrary alternative evolutionary models."
},
{
"claim": 2,
"assessment": "verified",
"evidence_tier": "literal_claim_experiment",
"claim_object_match": "exact",
"registered_system_executed": true,
"paper_or_released_scale": true,
"actual_model_or_dataset_used": true,
"destructive_control_executed": true,
"literal_claim": "Lemma 1 shows that repeated evolutionary selection induces conditional dependencies among variables that are absent under static selection models, so applying static-selection graphical models to evolutionary data can yield false causal discoveries (Section 2, Lemma 1).",
"paper_native_mechanism": "The executable conditions the unrolled graph on every earlier S variable, compares its generation-T d-separations with the static selection DAG, and runs causal-learn PC on evolved samples.",
"independent_oracle": "The custom Bayes-ball result is cross-checked against networkx.is_d_separator, and the finite-data consequence is tested against a paired no-selection SCM with shared SEM parameters.",
"not_proxy_reason": "The test uses the literal repeated-selection graph, static graph, conditioning sets, and selection-blind PC discovery consequence named by the claim.",
"native_scale_justification": "All 1,433,520 disjoint-set relations from 8,688 source models at T=1,2,3 are enumerated, plus twenty paired 5,000-sample PC trials at d=8.",
"source_locator": "arXiv 2606.05689v1, Section 2, Lemma 1 and the paragraph immediately following it.",
"independent_evidence": ["outputs/exact_graph_audit.json", "outputs/finite_native_algorithms.json", "outputs/results.json"],
"executed_outputs": ["outputs/exact_graph_audit.json", "outputs/finite_native_algorithms.json"],
"oracle_artifacts": ["outputs/oracle_gates.json", "SOURCE_EXCERPTS.md"],
"control_artifacts": ["outputs/destructive_controls.json", "outputs/finite_native_algorithms.json"],
"destructive_or_boundary_control": "Deleting all inheritance edges removes every one of the 17,712 converse-failure witnesses; switching selection off reduces native PC false adjacencies from 46 to 8.",
"result": "Lemma 1 has 0 violations in 1,433,520 exact relations and 17,712 strict converse-failure witnesses; paired PC finds 46 spurious adjacencies with evolution versus 8 without selection.",
"limitation": "Finite PC counts include ordinary type-I and type-II CI errors; the exact graphical enumeration supplies the decisive lemma result.",
"scope_boundary": "Verification concerns conditional-dependence structure and its false-discovery mechanism, not the frequency of false discoveries in every empirical domain."
},
{
"claim": 3,
"assessment": "verified",
"evidence_tier": "literal_claim_experiment",
"claim_object_match": "exact",
"registered_system_executed": true,
"paper_or_released_scale": true,
"actual_model_or_dataset_used": true,
"destructive_control_executed": true,
"literal_claim": "Theorem 1 proves that the clique-augmented DAG G^+ (Definition 2) fully captures all d-separation/conditional-independence constraints implied by the evolutionary selection model, without needing to explicitly model the selection variables (Section 3, Definition 2, Theorem 1).",
"paper_native_mechanism": "The executable forms Definition 2's topologically oriented ancestor-of-S clique and compares every tested d-separation in G^+ with the selected, fully unrolled G^(T).",
"independent_oracle": "A separately gated networkx d-separation implementation checks the custom Bayes-ball primitive on 11,984 random queries before the theorem enumeration is interpreted.",
"not_proxy_reason": "Both sides of Theorem 1's literal biconditional are evaluated directly; covariance, reachability, or a theorem restatement does not replace d-separation.",
"native_scale_justification": "The complete d=4 model space yields 1,433,520 exact biconditional cells over T=1,2,3, with a further 143,811 random cells through d=10 and T=5.",
"source_locator": "arXiv 2606.05689v1, Section 3.1, Definition 2 and Theorem 1.",
"independent_evidence": ["outputs/exact_graph_audit.json", "outputs/oracle_gates.json", "outputs/results.json"],
"executed_outputs": ["outputs/exact_graph_audit.json"],
"oracle_artifacts": ["outputs/oracle_gates.json", "SOURCE_EXCERPTS.md"],
"control_artifacts": ["outputs/destructive_controls.json"],
"destructive_or_boundary_control": "Replacing G^+ with the selection-blind S-deleted DAG creates 98,787 exact mismatches, so agreement cannot come from an inert augmentation.",
"result": "G^(T) and G^+ agree on all 1,433,520 exact d-separation relations with 0 violations; the naive S-deleted control mismatches 98,787 relations.",
"limitation": "Finite enumeration is exhaustive at four traits and randomized above that size; the theorem, not this audit, establishes unbounded-size validity.",
"scope_boundary": "Verification covers the paper's causally sufficient evolutionary DAG class and does not extend to dependent heritable-factor models."
},
{
"claim": 4,
"assessment": "verified",
"evidence_tier": "full_pipeline_reproduction",
"claim_object_match": "exact",
"registered_system_executed": true,
"paper_or_released_scale": true,
"actual_model_or_dataset_used": true,
"destructive_control_executed": true,
"literal_claim": "Theorem 2 establishes that applying standard constraint-based algorithms such as PC or GES (Algorithm 1) to G^+ is sound and complete: oriented edges correspond to true causal relations, while unoriented edges may reflect the presence of selection (Section 3, Theorem 2, Algorithm 1).",
"paper_native_mechanism": "The executable runs PC and GES from causal-learn on the paper's evolved SEM data and computes the CPDAG of each exact G^+ for the theorem's oracle-premise audit.",
"independent_oracle": "The local Meek closure is cross-checked on 600 models against causal-learn DAG-to-CPDAG, while alternative source DAGs are exhaustively searched for every unoriented edge.",
"not_proxy_reason": "The audit executes the named algorithms and evaluates the literal adjacency, oriented-edge, and alternative-DAG completeness clauses against their generating evolutionary models.",
"native_scale_justification": "All 8,688 four-trait source models cover 52,128 adjacency cells, 2,640 compelled orientations, and 39,030 unoriented-edge alternatives; PC/GES also run in 750 paper-scale trials.",
"source_locator": "arXiv 2606.05689v1, Section 3.2, Algorithm 1 and Theorem 2.",
"independent_evidence": ["outputs/exact_graph_audit.json", "outputs/oracle_gates.json", "outputs/synthetic_native_pipeline.json", "outputs/results.json"],
"executed_outputs": ["outputs/exact_graph_audit.json", "outputs/synthetic_native_pipeline.json"],
"oracle_artifacts": ["outputs/oracle_gates.json", "SOURCE_EXCERPTS.md"],
"control_artifacts": ["outputs/destructive_controls.json"],
"destructive_or_boundary_control": "A compelled CPDAG orientation is reversed in 328 eligible source models, and the Theorem 2 soundness checker rejects all 328 mutants.",
"result": "Across 8,688 exact models, 52,128/52,128 adjacency cells, 2,640/2,640 oriented edges, and 39,030/39,030 unoriented-edge completeness cases pass.",
"limitation": "The exact verdict assumes the theorem's large-sample faithfulness premise; finite-sample PC and GES can and do violate oracle-level orientation soundness.",
"scope_boundary": "Verification is for Algorithm 1 under Theorem 2's premises, not a guarantee that finite-sample CI testing always returns the population CPDAG."
},
{
"claim": 5,
"assessment": "verified",
"evidence_tier": "full_pipeline_reproduction",
"claim_object_match": "exact",
"registered_system_executed": true,
"paper_or_released_scale": true,
"actual_model_or_dataset_used": true,
"destructive_control_executed": true,
"literal_claim": "Theorem 4 shows that combining heterogeneous data from multiple environments/domains via the CDNOD-based procedure (Algorithm 2) improves identifiability of the evolutionary selection model compared to single-environment data (Section 4, Theorem 4, Algorithm 2).",
"paper_native_mechanism": "The executable builds Theorem 3's multi-domain G^{+I}, applies background-knowledge Meek closure, and runs causal-learn CDNOD on four domains with changed selection mechanisms.",
"independent_oracle": "Every multi-domain orientation is compared with the generating G and the complete single-domain orientation set; a separate finite run compares CDNOD with PC on the identical first domain.",
"not_proxy_reason": "The test uses the literal domain-index graph and the named CDNOD implementation on heterogeneous evolutionary data, not pooled PC or an edge-count formula.",
"native_scale_justification": "All 8,688 four-trait source models are crossed with 31 nonempty change sets for 269,328 exact configurations, plus twenty four-domain CDNOD runs with 5,000 samples per domain.",
"source_locator": "arXiv 2606.05689v1, Section 4, Theorem 3, Algorithm 2, and Theorem 4.",
"independent_evidence": ["outputs/exact_graph_audit.json", "outputs/finite_native_algorithms.json", "outputs/results.json"],
"executed_outputs": ["outputs/exact_graph_audit.json", "outputs/finite_native_algorithms.json"],
"oracle_artifacts": ["outputs/oracle_gates.json", "SOURCE_EXCERPTS.md"],
"control_artifacts": ["outputs/destructive_controls.json"],
"destructive_or_boundary_control": "Omitting Theorem 3's changed-selection ancestor expansion creates 583,356 orientation-soundness violations across the same 269,328 configurations.",
"result": "Multi-domain identification loses 0 single-domain orientations and is strictly better in 88,176/269,328 configurations; native CDNOD orients 4.65 correct edges versus 3.55 for single-domain PC.",
"limitation": "Strict improvement is model-dependent at 32.74%, while Theorem 4's universal statement is monotonic non-loss rather than strict gain in every configuration.",
"scope_boundary": "Verification assumes fixed causal and selection structures across domains with mechanism parameters changing as specified by the paper."
},
{
"claim": 6,
"assessment": "falsified_as_literally_registered",
"evidence_tier": "literal_benchmark_reproduction",
"claim_object_match": "literal",
"registered_system_executed": true,
"paper_or_released_scale": true,
"actual_model_or_dataset_used": true,
"destructive_control_executed": true,
"literal_claim": "The proposed identification procedure is validated on synthetic graphs of varying size and on seven real-world datasets spanning biology, agriculture, and social science (Section 5).",
"paper_native_mechanism": "The executable runs causal-learn PC and GES on 750 samples from the printed evolutionary SEM/reproduction pipeline and runs PC on the actual 626-by-8 PanTHERIA slice.",
"independent_oracle": "Generating DAGs provide exact synthetic causal adjacencies; Appendix D.2's seven integer tables are recomputed arithmetically, and PanTHERIA input bytes are independently pinned.",
"not_proxy_reason": "The named algorithms, paper-scale dimensions, generations, sample size, actual mammal dataset, and registered precision metric are used; oracle CPDAG precision is not substituted.",
"native_scale_justification": "The synthetic grid has d=10,15,20, T=1,2,3,5,10, N=5,000, and 50 seeds per cell; all seven real-data tables are audited and PanTHERIA is independently rerun.",
"source_locator": "arXiv 2606.05689v1, Sections 5.1-5.2, Figure 6, Appendix D.1, and Figures 13-19.",
"independent_evidence": ["outputs/synthetic_native_pipeline.json", "outputs/realdata_native_and_source_audit.json", "outputs/results.json"],
"executed_outputs": ["outputs/synthetic_native_pipeline.json", "outputs/realdata_native_and_source_audit.json"],
"oracle_artifacts": ["outputs/realdata_native_and_source_audit.json", "inputs/PanTHERIA_1-0_WR05_Aug2008.txt", "SOURCE_EXCERPTS.md"],
"control_artifacts": ["outputs/destructive_controls.json"],
"destructive_or_boundary_control": "Changing one printed DGRP numerator makes the arithmetic audit fail, and flipping one PanTHERIA input byte is independently detected by the pinned SHA-256.",
"result": "In 750 native synthetic runs, oriented-only PC beats the standard interpretation in only 3/15 grid-cell means and 0/5 d=20 cells; the PanTHERIA rerun gives 0.40 oriented versus 1.00 unoriented precision at the registered log/alpha-0.05 sensitivity cell.",
"limitation": "Only PanTHERIA is rerun from raw real data; the other six real datasets are audited from the paper's printed outputs, whose pseudo ground truth the paper itself calls an unreliable surrogate.",
"scope_boundary": "Falsification targets the compound empirical-validation claim and does not dispute that the authors ran analyses on seven named datasets or the exact oracle theorems in Claims 1-5."
}
]
}
|