ProCreations's picture
Publish validated ICML reproduction
177308a verified
|
Raw
History Blame Contribute Delete
1.88 kB
metadata
title: Reproduction - Causal Selection in Evolution
emoji: 📊
colorFrom: blue
colorTo: green
sdk: static
pinned: false
short_description: Exact graph audit and native pipeline rerun
tags:
  - trackio
  - open-reproductions
  - icml2026-repro
  - paper-mOcTXKawFY

Reproduction logbook

Local-only, semantic-v4 reproduction package for OpenReview mOcTXKawFY, “Causal Modeling of Selection in Evolution.” The source is pinned to arXiv 2606.05689v1.

Verdicts: Claims 1–5 are verified by exact enumeration and native algorithm runs. Claim 6 is falsified as literally registered because the paper-native 750-run finite-sample pipeline does not reproduce the claimed validation ordering: PC oriented-only precision wins only 3/15 grid cells and 0/5 cells at d=20.

Replay

python3 -m pip install -r requirements.txt
python3 reproduce.py --output-dir fresh_replay
python3 validate_evidence.py
python3 verify_manifest.py

The committed outputs/ and packaged_replay/ trees were generated in separate executions and are byte-identical. Runtime measurements are deliberately excluded from canonical evidence.

Evidence map

  • EVIDENCE_MATRIX.json: semantic-v4 claim contract.
  • pages/00-judge-evidence-scorecard/page.md: deterministic projection.
  • outputs/exact_graph_audit.json: exhaustive Claims 1–5 evidence.
  • outputs/oracle_gates.json: networkx and causal-learn cross-checks.
  • outputs/finite_native_algorithms.json: paired PC and CDNOD executions.
  • outputs/synthetic_native_pipeline.json: paper-scale PC/GES experiment.
  • outputs/realdata_native_and_source_audit.json: raw PanTHERIA rerun and seven-table audit.
  • outputs/destructive_controls.json: one executed break-test per claim.
  • BUNDLE_SHA256SUMS.txt: recursive package manifest.

No publication or remote mutation is performed by any script in this package.