Overview & result
Yijun Yu shows jev-rs, built on Jev: A Rust, Jev-compatible /v1/systemone engine, one prefill per judgement.
From the original post
jev-rs by @yijunyu: A Rust, Jev-compatible /v1/systemone engine, one prefill per judgement.
How Jev fits in the loop
- The state is rendered once as a shared prefix; each question is a suffix with options labelled A, B, C.
- The backend returns raw next-token log-probabilities for the labels; nothing is decoded.
- A restricted softmax over the labels is divided by a fitted temperature per question type and option count.
- The result is served on a TypeSafe-compatible endpoint and as an MCP tool.
How to reproduce
- Review the source documentation and repository at https://github.com/yijunyu/jev-rs.
- Configure environment with TypeSafe Jev / System One API keys.
- Run test prompts against the decision endpoints.
Why this build matters
Shows Jev in practical use for developer tools and apps.
Limitations
- At most 26 options per question; Jev accepts 255.
- Zero-shot only, with no Jev-style training.
- Scorer backends today are llama-server and OpenAI-compatible hosts.