LLM evals

AI systemsDefined term

An eval is a repeatable measurement of model behavior: a dataset of cases, a task definition, and a scoring method — exact match, rubric, or a judging model. For agents, evals extend to whole trajectories: did the loop reach the goal, at what cost, with what failure modes.

Evals are the difference between a demo and a system. Because model updates shift behavior invisibly, teams pin expectations in eval suites the way software teams pin regressions in tests — and judge-model evals inherit their own calibration problem, which is why humans audit the judges.

Browse the glossary

AI systems