ZenML docs: judge agent sessions with TypeSafe's jev model, ~270ms per call

strickvl · x · 2026-09-24

ZenML published full docs for Kitaru judge evaluations: TypeSafe's jev takes JSON state and typed questions, returning yes/no probabilities, a chosen label with confidence, or a position on ordered levels, one result per question.

The docs compare three approaches — deterministic evaluators (free, code rules only), the typed model judge (270ms per call in exploratory runs, can catch invented facts), and hand-written LLM judges (flexible but cost/repeatability depend on model and prompt) — and when to use each.

Original post →

More from coding & agent

coding & agent channel →