Arize AX Adds Native Jev-as-a-Judge Integration for Production Trace Evaluation
Arize AI integrated TypeSafe’s Jev directly into Arize AX, reporting threshold-tuned Jev matched Claude Opus 5 at 87% hallucination-detection accuracy at 1/300 cost and 23x speed.
AI observability platform Arize AI became one of the first production developer-tool vendors to ship a native Jev integration in late September, introducing “Jev-as-a-Judge” directly inside Arize AX so teams can evaluate live production traces with boolean, choice, and rubric-score questions.
Matching frontier LLM judges at 1/300 the cost
Writing on the Arize blog, Chris Cooning shared internal benchmark results testing Jev on hallucination detection against full LLM judges. Once tuned with an appropriate probability threshold, Jev matched Claude Opus 5 at 87% accuracy while running at roughly 1/300 of the cost and 23x the speed in Arize’s test.
Enabling 100% production trace coverage
Historically, teams running LLM-as-a-Judge pipelines had to sample a tiny slice of live traffic—often 1% to 5%—because grading every conversation with a frontier chat model cost as much as serving the product itself. Arize highlighted that cutting evaluation cost by two orders of magnitude changes operational practice: teams can score far more production traffic automatically, surface rare agent failure modes faster, and reserve slower generative or agentic judges only for cases where a written explanation or deeper investigation is needed.