# Contributing

Contributions are welcome when they preserve the evidence boundary.

1. Register the exact source data and checksum.
2. Freeze the operator, score, target, and controls before inspection of the result.
3. Report positive, neutral, and negative outcomes.
4. Separate model-product diagnostics from observational likelihoods.
5. Do not count reused samples or overlapping channels as independent confirmation.
6. Add a deterministic reproduction test where practical.
