A CLI that takes the output of grep, git ls-files or a test run and returns each candidate line with a calibrated score, sorted, with an exact total and a cut line you read above.
User repository on GitHub
Don't burn your model's context on a search flood — pipe it in and read above the cut line: one calibrated number per candidate, ~30 rows whether the pool is 40 or 4,000. It doesn't search; it decides what each hit is.
Repo stats from the GitHub API, cached Oct 2026.
A search comes back as hundreds of lines: too many to read, too risky to guess at. Pipe the candidates in — raw grep -rn, git ls-files, a test run, no reformatting — and each one comes back with a calibrated number, sorted, with an exact total and a cut line. The reply is 30 rows whether the pool was 40 or 4,000. Built on the three typed judgments of TypeSafe Jev: noul (is this hit real?), choice (what is it?), score (how bad is it?).
Quoted from the project's README. Read the full README.
Watch out
Each call ships your candidate lines and your question to the Jev endpoint, and the README's instruction is blunt: if the repo must not leave the machine, do not point jev at it. Verdicts are cached in ~/.cache/jev/cache.json keyed by question and line, which is only sound because Jev is self-consistent by design, and --fresh recomputes. It is explicitly not for retrieval, for pools small enough to read, or for anything a compiler, linter or test already decides. Copy reviewed against the repository README on 2026-10-04, not from running it.
E-FL/typesafe-as-a-judgeJavaScript
LXBWOW/dsh-context-curatorJavaScript
xienda/dsh-jev-verifyJavaScript
PyModel/jev-judge-mcpPython
mizchi/jev-test-filterTypeScriptPosts from builders shipping with Jev right now.
