docs/EVALUATION.md

# Evaluation resources and benchmark governance

Evaluation data requires the same provenance discipline as training data plus stronger controls
against contamination and answer leakage.

## Resource record

For each benchmark or evaluation set, record:

- canonical repository, paper, version, revision, and licence;
- intended task, population, language, and domain;
- metric definitions, aggregation, uncertainty, and known limitations;
- access or gating terms and whether examples may be redistributed;
- contamination risk and separation controls;
- prompt, decoding, scoring, and harness versions;
- sample exclusions and failure-handling policy.

## Evaluation rules

- Keep evaluation inputs out of training and prompt-development mixtures.
- Do not publish gated or leakage-sensitive examples in screenshots, logs, or documentation.
- Pin benchmark revisions and harness code.
- Prefer deterministic or independently auditable metrics.
- Report confidence intervals or run-to-run variation when stochasticity matters.
- Separate capability, safety, robustness, efficiency, and user-experience claims.
- Record failures and invalid runs; never silently discard them.
- Treat model-graded metrics as measurements from a named, versioned judge, not objective truth.

## Example evaluation families

The research Collections nominate real external resources for study. Inclusion in a Collection is
not permission to copy, train on, redistribute, or reveal their contents. Their current licences,
gates, cards, and contamination guidance must be reviewed at the time of use.