bsafronov Evaluate Skill
Evaluate behavior without changing the Skill.
- Inspect the Skill, its cases, and descriptions of nearby Skills that could compete with it.
- Exercise realistic selection, rejection, and outcome cases. Exercise composition only where multiple Skills should materially help.
- Keep expected answers hidden from an independent evaluator when independent execution is available.
- Record source revision, harness, model, result, variance, and concise evidence.
- Report failures by observable behavior: wrong selection, boundary violation, harmful composition, or unmet outcome.
For Promotion evidence, run every required case twice on the reference harness and once on each required compatibility harness. Stop with a read-only Evaluation Report; do not edit or promote the Skill.