Contributing

Contributions are welcome for runtime portability, deterministic validation, documentation, benchmark reproducibility, and the sandboxed demo.

Before opening a pull request:

keep model proposals separate from external authorization and execution;
add a focused regression test for behavior changes;
run pytest -q, ruff check ., ruff format --check ., and python scripts/verify_public_release.py;
do not add raw datasets, official or private labels, model checkpoints, credentials, internal infrastructure records, or unverified benchmark claims; and
preserve all upstream licenses and attribution.

Benchmark changes must identify the dataset revision, split construction, metric, evaluator, decoding settings, model revision, and sample-level evidence needed to reproduce the claim.