Reproduction instructions
uv run --with numpy==2.2.6 --with matplotlib==3.10.3 reproduce.py
(git clone https://github.com/mmschlk/shapiq.git vendor_shapiq && cd vendor_shapiq && git checkout 12ec2878fb61b94c01a8c1260a18108eac165538)
# Download dataset-latest.zip from the pinned TALENT revision linked below,
# extract its data/ directory as talent_data/data/, then run:
uv run --python 3.12 --with ./vendor_shapiq --with causal-learn claim56_reproduction.py
(cd results && sha256sum -c SHA256SUMS)
(cd results/claim56_full && sha256sum -c SHA256SUMS)
The deterministic CPU run uses seed 18355 and fails closed on paper/source
checksum drift. It independently implements the published pseudocode and
produces machine-readable exact-graph, boundary-sampling, identifiability, and
complexity audits.
The named-estimator rerun uses the public shapiq repository pinned above. The real-data rerun uses the public TALENT release pinned at revision 7bd276bcb7f6b4c0998025855528bd76bd88f13d: https://huggingface.co/datasets/LAMDA-Tabular/TALENT/tree/7bd276bcb7f6b4c0998025855528bd76bd88f13d
The 648 paired Claim-5 runs and all 156 Claim-6 dataset rows are preserved in
results/claim56_full/. The TALENT archive itself is not duplicated in this
bundle.