Review pull requests with calibrated bug risk

01 · Role in the system

What Jev does here

The tool diffs a branch or commit range, clusters changed hunks, and asks Jev one call per cluster: a Noul gates whether each suspected issue is real at all, Score questions rate severity, and the PR level gets an overall risk score plus a needs_human_review Noul - the sample run reports 2.1/3 risk with needs-human at 0.93. Every run appends raw scores to a metrics JSONL so judgments can be tracked over time, and a --golden mode checks the judge against labelled cases before you trust it. Output is advisory and labelled as such: findings carry their own is_real probability next to the hunk, so a reviewer can see how sure the judge was. Because the model only returns typed answers, the plumbing - git, thresholds at 0.50, report formatting - stays deterministic Python in one file.

02 · Control boundary

Where Jev sits

Jev as the judgment layer of a deterministic review pipeline: code clusters hunks and formats reports, one call per cluster returns real-issue Nouls, severity Scores and a needs-human probability, and a metrics log plus golden evals keep the judge accountable.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• Self-described experimental and advisory only; findings are suggestions, not verdicts.
• Single-maintainer tool with no hosted service; run locally with your own API key.
• The 0.50 real-issue threshold is a fixed author choice; no benchmark against human reviewers yet.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• equationalapplications ↗Community · github · public · checked 2026-09-28