The option applies to simulations, scheduled tests, and production monitoring.
Jev accepts conversation state, tool calls, and traces together with typed questions. It returns bounded results such as pass/fail, yes/no, an enumerated choice, or a rubric score, with probabilities and usually a confidence value. In Bluejay, it is available for pass/fail, yes/no, enum, and quantitative metrics on the text route; qualitative and JSON metrics, plus audio-route metrics, continue to use other judges.
Each Jev metric can have a confidence threshold. Results at or above that threshold stand; lower-confidence checks are re-graded by an LLM judge, while the product stores one result per check. Bluejay manages model access, so users do not supply a Jev API key. The company says Jev responds in well under a second and cites vendor benchmark ranges of 20 to 200 times faster and 40 to 400 times cheaper than frontier models. The source does not independently establish those performance or cost claims, or Jev's accuracy on any customer's data.
https://getbluejay.ai/blog/jev-as-a-judge