Run Skill Eval

Run the Skill benchmark and return its verdict. Output cases grade generated behavior; should_not_trigger cases use the production semantic Skill router and become inconclusive when semantic evidence is unavailable rather than falling back to lexical matching. A passing routing negative is a safety gate and cannot by itself prove that a Skill helps on intended work. strict=true runs the output cases through the median-of-3 with/without gate; the streaming variant is /eval/stream.

Path Parameters

Query Parameters

Response Body