Testing Jev as a confidence-scored router to offload small LLM calls in oncology agents

Ubunta · x · 2026-09-26

The author has been trying Jev this week, likening it to a Bloom filter that runs a cheap check before expensive work — except Jev outputs a confidence score rather than a hard guarantee.

Working with real-world oncology data, their AI agent plans and runs analyses on a data warehouse, and the big model wastes time on trivial calls like counts or survival analyses. They are testing Jev for three roles: routing, ask-or-act decisions, and plan checks.

The architecture keeps reasoning in the large model while data access rules stay deterministic. The open question is whether Jev's confidence holds up on clinical data.

Original post →

More from coding & agent

coding & agent channel →