Overview & result
TianyuCodings shows JevHarness, built on Jev: LLM-written Jev harnesses, with reward reflection over the full trajectory.
From the original post
JevHarness by @tianyucodings: LLM-written Jev harnesses, with reward reflection over the full trajectory.
How Jev fits in the loop
- You describe a task to Codex or Claude Code with the JevHarness skill installed.
- The authoring LLM writes a harness: feature construction, Jev questions and criteria, and the logic that turns answers into actions.
- A task adapter owns observations, legal actions, side effects and scoring, so the harness cannot rewrite its own reward.
- Once frozen, the harness runs on its code and Jev calls alone; rewards and execution traces can optionally be used to improve it.
How to reproduce
- Review the source documentation and repository at https://github.com/TianyuCodings/JevHarness.
- 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 browser and autonomous agents.
Limitations
- The README says the reported improvement is an example result on the selection eval set, not an independent estimate on unseen games.
- Frozen harnesses still need Jev if they contain Jev nodes.
- A hosted model alias does not pin future provider behaviour, so fresh calls may not reproduce stored ones.