NotJev

Checked, not fully reviewed. We ran its 187 tests (all pass) and drove notjev serve against a stub backend whose log-probabilities we chose. No Jev calls.

NotJev presents your options as A., B., … asks the backend for one token with logprobs, and keeps only the letter mass. The arithmetic is right: with letters at 0.70 and 0.10 and the rest on another token, it answered 0.875; it asked for max_tokens: 1, top_logprobs: 20, temperature: 0. Every answer carries a notjev block with coverage — how much mass landed on the options at all — which is the number to watch: at 10% coverage it still answered, with a low confidence and degraded: false.

The SDK now reads it

When we first checked it, /v1/systemone grouped answers under nouls, choices and scores, so the official SDK returned without error and found nothing. We reported it; the author fixed it within a day in v0.2.3, and fixed two things we had not reported: score legend and probabilities are now keyed by level ("0", "1", …), and GET /v1/models returns the { models: [...] } shape. We reran our script with a noul, a choice and a score through @typesafe-ai/sdk 0.6.0: all three answers came back under their question names, probabilities[1] was readable, and client.models.list() worked.

This is a protocol change: code that read answers.nouls from the old route needs updating. The native /v1/decide route is unchanged.

Its confidence, 1 − H(p)/ln K, is not the formula TypeSafe published on 3 October, and the README now says so, so the two are not comparable. Some requests are refused with a 422 rather than answered: null instructions, a state that is not a string, and a choice with more than 26 options.

Apache-2.0, 21 stars.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Oct 5, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.