Contributions should improve observable learning decisions rather than add theory names or output sections.
- Describe a realistic failure or missing behavior.
- Add or update a behavioral case in
evals/cases.json. - Make the narrowest instruction or implementation change that addresses it.
- Add executable tests for deterministic behavior.
- Run the unit tests and skill validator.
Avoid universal rules based on one anecdote. Preserve user agency, project scope, privacy, and the distinction between capability evidence and review stability.