Why add a second judge instead of trusting Claude alone?

Claude still proposes candidates, but Jev rates each one with independent probabilities that code turns into keep or drop verdicts. The digest shows a real bug kept, a false positive kept near threshold, and a guarded case correctly dropped.

What do the five MCP tools actually ask Jev?

Yes: review checks real, reachable, guarded, and severity; debugging weights explains and supported; planning compares best, safest, and simplest plus risk; search filters grep hits before reading them.

How do I read Noul versus Choice and Score answers?

A Noul is a yes probability with no confidence field, while Choice and Score add confidence as distribution concentration rather than correctness. Values near 0.5 or flat splits usually mean the question was ambiguous.

What is the setup catch with keys, limits, and fan-out cost?

Needs Node plus a recent Claude Code, installs from the marketplace with no build step, and takes the key from plugin option, settings file, or environment. Source reading caps bytes and sources, and oversized states error rather than truncate.

What does the plugin actually read off disk, and what does it refuse?

The FsSourceReader resolves each path against the session's MCP roots, realpaths it, re-checks containment, refuses secrets and key material, slices the requested line range and caps the bytes. The state it sends is the task, the preset fields and those sources. Token counts are estimated from character counts and deliberately rounded up, since there is no tokenizer in the loop, and nothing is silently dropped.

What are the configurable limits on a call?

Source reading is on by default; turn it off and only text passed in the call is sent. Max bytes per source is 65536, longer snippets are truncated and say so in the snippet. Max sources per call is 20, and max requests per call is 8 so a wide question set cannot run away. The model defaults to jev-latest, and pinning jev-1.13.0 keeps the tuned thresholds stable across releases.

What does a judgement actually cost in money and time?

Input tokens are billed at $0.042 per million and answers come back in roughly 150 ms, which is what makes asking on every review viable. A real run against this repository's own source reader was one request, about 1714 tokens of state and 2616 input tokens, keeping two of three candidate findings. The cost is attached to the compact verdict block each tool call returns.

CommunityJev Practices & Patterns

Claude Jev — buchmark

Claude Code plugin that routes code review, debugging, planning, and search triage through typed Jev judgements with fixed thresholds applied in code.

See highlights

Open github.com

Stars: 7
Language: TypeScript
License: MIT
Published by: Community
Author: buchmark
Last verified: Oct 2026

buchmark/claude-jev

User repository on GitHub

View on GitHub

Claude Code plugin that scores review findings, debug hypotheses and design options with TypeSafe's Jev — calibrated probabilities instead of one more opinion.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is Claude Jev — 0x7067.

About this project

A Claude Code plugin that turns a question about your code into a typed TypeSafe Jev query and hands back calibrated probabilities. Jev is a System One model. It does not write text, code, or explanations. It takes a state plus named typed questions and returns numbers. That makes it a second opinion an agent cannot bluff its way past: the answer is a probability, and it either clears your threshold or it does not.

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Adds an independent calibrated judge to review findings, hypotheses, options, and hits.
Keeps question texts and verdict thresholds in one catalog directory for human review.
Estimates tokens, batches within state limits, runs concurrently, and merges answers.
Refuses oversized states instead of truncating and reports verdicts with costs attached.

Watch out

Probabilities near thresholds are weak signals to read yourself, and wide fan-outs resend the state with every request. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

Kev

Self-hosted Apache-2.0 decision engine with a Jev-style wire format that turns a model you point at it into typed choice, score, and noul answers.