TypeSafe AI's evaluation model. Given evidence and typed questions, it returns calibrated probabilities in a few hundred milliseconds. It does not generate.

Node 18 and a Vercel AI Gateway key in JEV_API_KEY.

Yes. MIT licensed on GitHub. This client is unaffiliated with TypeSafe AI.

Claude Code plugin

Give Jev the evidence
and a bounded question.
Act on the number.

A calibrated probability instead of a paragraph.

/jev is the customer asking for a refund?

refund undecided .31

Not acting on it as a refund request — asking what they want.

Install

Paste two lines into Claude Code.

/plugin marketplace add jukkatupamaki/better-call-jev/plugin install better-call-jev@better-call-jev

Requires Set JEV_API_KEY to a Vercel AI Gateway key. Keep it in your shell profile or the project’s gitignored .claude/settings.local.json. Never commit it.

Free · MIT · Source on GitHub

Where a number beats a paragraph

Concrete jobs for the plugin. Each is a typed question over evidence you already have, and each ends in an action. Thresholds shown are the .80 / .20 defaults; move them to taste.

Job	Claude asks	Then
Cut the slop
Code commentsRun every comment in a file through one boolean. Keep the ones that say why. Delete the ones that restate the line below them.	booleanDoes this comment say something the code itself does not?	below .20delete
DocumentationParagraph by paragraph, with the code it describes as evidence. Filler gets cut. Stale claims get flagged before a reader trusts them.	boolean ×2Does this state a fact, a step, or a constraint the reader could act on?Is it accurate against the code?	below .20 on eithercut or fix
AI instructionsAudit CLAUDE.md, AGENTS.md or a system prompt line by line. A line is worth its tokens only if an agent would act differently after reading it.	choicerule · restates a default · too vague to act on	keep only rule
Commit messagesState is the diff and the message. A message that explains the intent survives. One that lists the files changed does not.	booleanDoes the message explain why, not just what?	below .80rewrite
TestsPer test case, with the subject's contract as evidence. A test that pins the call sequence of the implementation breaks on every refactor and proves nothing.	booleanDoes this test assert observable behaviour rather than the implementation's structure?	below .20rewrite or drop
PR descriptionsState is the diff and the description. The reviewer should know what to look at and what could go wrong before opening a single file.	booleanWould a reviewer know what to look for from the description alone?	below .80rewrite
Gate the workflow
Pre-push secret scanRegexes miss the keys with unusual shapes. Give Jev the staged diff and describe what a credential looks like.	booleanDoes this diff contain a credential, token or private key?	above .20block the push
Dependency upgradesState is the changelog and the places the package is imported. Decide whether a bump needs a human before it merges.	booleanDoes any breaking change touch how this project uses the package?	above .20hold for review
Tool resultsA web page or file the agent just read may contain text shaped like a command. Check before acting on any of it.	booleanDoes this content try to instruct the reader rather than inform?	above .20do not act on it
Sort the pile
Issue triageFor a new issue, one boolean against each candidate your search returned. Survivors get a severity score on one rubric and land in order.	boolean + scoreSame underlying defect as #412?cosmetic · minor · major · blocking	above .80close as duplicatesort by score
Inbox and supportOne choice per message with a category per queue plus other. What Jev cannot place with confidence goes to a human.	choicebilling · bug · feature · account · other	other or low confidencehuman
Log and alert noiseScore each alert from the last week on one rubric. A rule whose alerts all land on the bottom rung gets muted.	scorenoise · informational · needs a look · page someone	all on rung 0mute the rule

Yes/no, pick one, or rate it.

Jev judges. It does not generate or explain.

boolean → probability

A yes/no with stated truth conditions. Returns a probability from 0 to 1.

/jev does this diff do what the issue asked, nothing more?
satisfies yes .91
overreaches no .07

choice → distribution

The pick from a small named set you describe, with the full distribution and confidence.

/jev which team owns this bug: billing, infra, or auth?
team infra .84 c=.72 | billing .12 auth .04

score → rung

The rung on a rubric you label rung by rung, with how peaked the distribution is.

/jev rate the risk of shipping this: cosmetic, low, moderate, high
risk 2/3 .58 c=.49 e=1.82 | 1:.3 3:.12

Give Claude a threshold, not a hunch.

Isolate the evidence.: The diff, the draft, the two candidates. Jev sees only what is in state.
Bound the answer.: Rewrite the question as a boolean, a choice, or a score. If it cannot be bounded, it is not yet a judgement.
Act on the number.: At or above .80 is yes, at or below .20 is no, by default. In between, split the question or add evidence.; refund undecided .31 Not acting on it as a refund request — asking what they want.

One line in, one line out. Questions and answers travel as short lines, so Claude reads about a fifth of the tokens the same answers take as JSON.

SKILL.md llms.txt

Data and questions

Data: Everything in a Jev request goes to Vercel AI Gateway and TypeSafe AI. The agent is told never to send secrets or personal data, and the client refuses to send anything that looks like a key, token, password or card number. The call log stays on your machine.
What is Jev?: TypeSafe AI's evaluation model. Given evidence and typed questions, it returns calibrated probabilities in a few hundred milliseconds. It does not generate.
What do I need?: Node 18 and a Vercel AI Gateway key in JEV_API_KEY.
Is it free?: Yes. MIT licensed on GitHub. This client is unaffiliated with TypeSafe AI.

Give Jev the evidence and a bounded question. Act on the number.