full-audit-v2/source_audit_v2.md

# Source and scope audit

## Paper identity

- Paper: *Causal Modeling of Selection in Evolution*
- arXiv: https://arxiv.org/abs/2606.05689
- OpenReview: https://openreview.net/forum?id=mOcTXKawFY
- Audited PDF SHA-256: `5887ccd5ca9e175b1c80f1169d18fb3759ab9e606861c823244d9c3b5efea1c9`

## Paper anchors used

- Definition 1: the time-unrolled evolutionary-selection DAG and its four edge families.
- Definition 2 and Theorem 1: clique augmentation over selection ancestors and equality of the relevant conditional-independence model.
- Algorithm 1 and Theorem 2: interpret the CPDAG returned by PC/GES; oriented edges must be sound under the stated large-sample assumptions.
- Algorithm 2 and Theorem 4: add the exogenous domain index and use multi-domain information to identify additional directions.
- Section 5 and Appendix D.1: ER-DAG synthetic generator, average degree 2, `d/5` selection parents, coefficient magnitudes 0.5–2, noise variances 1–4, six offspring bins, `N=5000`, PC alpha 0.05, and GES L0 penalty 2.
- Appendix D.3: 2024 PUMS and Fisher-Z PC real-data analysis.

The scripts are independent clean-room implementations. No author repository was located or used.

## External sources

- causal-learn: https://github.com/py-why/causal-learn
- NetworkX: https://github.com/networkx/networkx
- 2024 ACS PUMS documentation: https://www.census.gov/programs-surveys/acs/microdata.html
- California person ZIP: https://www2.census.gov/programs-surveys/acs/data/pums/2024/1-Year/csv_pca.zip
- Downloaded ZIP size: 70,114,146 bytes
- Downloaded ZIP SHA-256: `fa3e473e13b09bb6d9f915c7f5c58b8405a8e48d22c43f30d43055a8a6c52882`

## Reproduction boundary

The oracle audit verifies finite sampled instances of the graph-theoretic statements, not their universal proofs. The discovery audit uses paper-native dimensions, generations, sample size, distributions, PC alpha, and GES penalty, but is not author-code execution and may differ in implementation details or random seeds. The real-data run covers California PUMS only; it does not cover the paper's remaining 49 states, DGRP, Cranial, Panzea, PanTHERIA, AVONET, or CSES. PUMS lacks causal ground truth, so it cannot measure edge precision.