The review queue does not auto-approve
jev_review_queue.py explicitly orders items for manual review without an acceptance threshold. Its probabilities are observations, not verified facts.
Inspect the supporting source ↗REPOSITORY ANALYSIS
azdaja
Bare, open-source RLM layer for existing coding agents
E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
- Stars
- 13
- Forks
- 0
- License
- MIT
- Last commit
- 2026-09-19
What Jev does here
Jev receives selected source material and explicit questions. It can rank passages, classify records, assess whether evidence supports a claim and judge whether records match for semantic joins. Its returned probabilities let the RLM compare results and decide what to inspect, combine or explore next.
Keeps complete sources locally and binds semantic observations to hashes and byte windows for manual review.
Fixed-version documentation and implementation material were inspected on 2026-10-24. No recursive model or batch inference was run. The project’s separate benchmark results were not reproduced in this review.
What task does this project fit?
Keep full source material locally while sending bounded semantic windows for manual review.
Material analysed：
From input to action
- The evaluator holds original inputs and performs exact parsing or filtering in code.
- The batch example creates hashed source windows with byte ranges and a Noul match question.
- Stored results feed a review queue that distinguishes observations, failures and ambiguous in-flight work.
Implementation details that matter
Provenance survives windowing
jev_batch_review.py binds source SHA-256, byte ranges and the question to requests. The reader can check the exact window rather than trusting an untraceable summary.
Inspect the supporting source ↗When to choose it
Choose this pattern when source size exceeds a convenient prompt and evidence must remain recoverable. It is a recursive/evaluator layer, not a standalone Jev replacement or agent harness.
How to inspect it yourself
Read _request and prepare in jev_batch_review.py, then the review queue’s validation. Verify that the same source hash and byte window identify each observation and that unknown/in-flight entries are not treated as accepted matches.
Limits of this review
- No recursive model or batch inference was run.
- The project’s separate benchmark results were not reproduced in this review.