CONTRIBUTING.md

Contributing to referent

Bug reports should include a minimal data-generating example, the intended predictive estimand (conditional or total), package versions, and whether the data are training, calibration, or held-out evaluation rows. Never include identifiable participant data.

For code changes:

1. Add a regression test that states the statistical contract or invariant.
2. Run devtools::test() and R CMD check --as-cran under a UTF-8 locale.
3. For numerical changes, include an independent oracle, metamorphic law, or adversarial counterexample; do not update a tolerance only to make a result pass.
4. Keep PCNtoolkit evidence classes and uncertainty estimands explicit. A better held-out score never excuses a conformance or calibration failure.
5. Do not commit raw external cohort data. Commit only small result tables and receipts with source URLs and hashes.

Pull requests that alter public functions, bundle schemas, comparison gates, or evidence classifications should update the relevant vignette and NEWS.