Jev technical blog and early access
The end of Diogo’s launch thread links to the official technical blog and early-access entry point for further reading.
Testing Jev as an LLM Judge
A Japanese developer tested Jev as an LLM judge with 72 misleading wrong answers, finding it reliably catches fully wrong answers but may miss partial errors. It runs at ~0.2s per case, ~2 cents per 1,000 cases, and they are also trying to route only high-risk samples to an LLM judge.
Jev は LLM ジャッジに使えるか、紛らわしい誤答72件で確かめました。 ・回答全体が誤っているものは確実に見抜く ・一部だけ誤っている回答は見逃すことがある ・1件約0.2秒、1,000件で約2セント 見逃した判定は確率で見分けられたので、そこだけ LLM ジャッジに回す使い方も試しています。
· 4 likesOpen on X
More like this
Getting started with Jev: waitlist and Agent Skill
A community guide walks developers through requesting access and installing the TypeSafe Agent Skill.
Reverse-engineered a Jev-like architecture with open-source training repo
The author says they reverse-engineered a Jev-like architecture from its type and shares a GitHub repo for training your own jevlikes. Link
