AI Operations · Glossary
Per-Turn Evaluation
Per-Turn Evaluation is evaluation that runs at the granularity of a single conversational turn, not at session end, providing the unit-level grain of EvalOps.
Per-Turn Evaluation matches the granularity of decisioning. An Interactive Agent makes decisions turn by turn; evaluation that aggregates only at session end loses the locus of the behaviour it scores. Per-Turn Evaluation runs criteria against each turn's input, decision, and output. It supports LLM-as-a-Judge scoring, deterministic checks, and policy-adherence verification at the same grain the trace records. The output feeds Evaluation Suites and the autonomous improvement loop. Per-Turn Evaluation is what lets Co-Pilot surface patterns at the turn level when drafting refinement proposals.