Evaluation Principles
Operational principles for building and reviewing AI interaction safety evaluations.
1. Define The Harm Model
Describe the behavior being tested before writing detectors. Separate direct harm, unsafe advice, dependency language, false continuity claims, and sycophancy.
2. Keep Benign Controls
Every risky fixture should have benign controls. This prevents detectors from blocking normal support, education, or fictional content where appropriate.
3. Measure Failures
Report false positives and false negatives. A tool that only reports wins is not ready for serious safety work.
4. Preserve User Agency
Safety responses should provide options, encourage qualified help where relevant, and avoid replacing user judgment with model authority.
5. Avoid Unsupported Authority
Do not market a detector as certified, clinical, legal, regulatory, or production-grade without qualified independent review.
6. Use Public Taxonomies Carefully
MITRE ATLAS, OWASP GenAI, and NIST AI RMF are useful references. Mapping to them is not the same as certification.
7. Version Everything
Version fixtures, detector configuration, package builds, and published results so other reviewers can reproduce the claim.