# Verification record

Verified locally with Python 3.10.20, NumPy 2.2.6, pandas 2.3.3, SciPy/scikit-learn from the `oa_tl` environment, and pyfaidx 0.9.0.4.

## Package integrity

`verify_dataset.py` passed:

- all package SHA-256 entries matched;
- 168 supervised rows, 56 positives and 112 controls;
- fold row counts `39 / 30 / 45 / 27 / 27`;
- 56 complete one-positive/two-control matched triplets;
- all supervised and pretraining sequences are 3,000 bp REF/ALT pairs with one centered substitution;
- 20,000 unique ordered pretraining examples were present.

## Source and feature reproduction

`verify_reproduction.py` passed:

- the governed 330-row development input matched SHA-256 `f3d29d9397509f7162d31bb9cc3f9e41e08e2507bceb14123a9a05c988017eb0`;
- source selection reproduced the exact 168-row OA2025 PIP≥0.30 cohort;
- the full-graph five-fold assignment matched SHA-256 `b0acffb900e897160599de1269fe3b27e70cff90bd37c557417a2ffa86389da4`;
- exported example IDs, labels, paired sequences, and outer folds matched the reconstructed cohort row-for-row;
- recomputing the 20,000 × 9,216 sparse feature matrix produced CSR SHA-256 `a9ce199b9a6e7c6185cd3945ef468e0051f89fc813277e86ab4c8ce23cdbc858`, identical to the original pretraining receipt.

These checks establish artifact identity and reproducibility. They do not add external biological validation or clinical validity.