Overview & result
Super Jev places one judge between an agent and your data. The agent asks in its own words; Jev finds the file, checks the claim, permits the action and remembers what you approved, while the agent stays responsible for the answer. Per the README, the cache fills only from your own approvals, and Jev handles yes/no and which-one questions while the LLM keeps the writing.
From the original post
A small, extensible decision-to-action harness for TypeSafe Jev
How Jev fits in the loop
- The agent asks a question in its own words.
- Jev ranks the connected files, and the agent reads the top file itself and answers.
- You approve the result, mark a miss or add a fact that has no file.
- Repeated wins for the same file on the same question are saved after a content check, so the next ask returns the list at once.
How to reproduce
- Inspect the original showcase and source material at https://github.com/Kevthetech143/super-jev
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
The cache fills only from your own approvals, and the agent stays responsible for the final answer.
Limitations
- Cache hits match exact wording only, not paraphrases.
- There is no automatic sync: data comes in through connectors you run.
- It does not answer unattended or browse the web.