Jev turns LLM decisions into multiple choice: no output tokens, just confidence scores

femke_plantinga · x · 2026-10-14

TypeSafe AI launched Jev on September 15, and OpenAI opened up its own version last week, quickly turning classification-only decision models into their own category. Unlike an LLM that writes answers token by token, Jev only does multiple choice: you give context, a question, and the allowed answers, and it scores each one.

Using the same "I was charged twice" support ticket: the LLM path needs a prompt asking for JSON, writes its answer token by token, and your code parses it and retries on format breaks. Jev instead outputs scores per allowed answer (e.g. billing 0.91), and your code routes above a threshold or escalates to a human below it.

The score is the key difference: LLMs sound equally confident when guessing, while Jev tells you how sure it is, letting you decide where a human steps in—while avoiding output-token costs and brittle JSON parsing.

Two caveats: TypeSafe hasn't disclosed how Jev works internally, so this is an input/output comparison; and a valid answer can still be the wrong team, so test on your own tickets first.

Original post →

Jev turns LLM decisions into multiple choice: no output tokens, just confidence scores

More from coding & agent