How is Jev different from just asking a chat model to decide?

A chat model generates text. Jev takes a state plus typed questions and returns calibrated probabilities in one fast round trip, and your code thresholds the numbers to decide.

Do ten questions cost ten times as much as one question?

Jev reads the state once and answers every question at the same time, so speculative fan-out such as eight ticket priorities plus an incident flag costs about one request.

What do I do with the confidence value besides logging it?

Choice and Score return both the answer and a confidence signal from the distribution shape, so confident answers can act automatically and unsure ones route to a human.

Which runnable examples does the playground actually include?

It walks through spam classification, a market-data call, an agent tool-call guardrail, and multi-ticket triage, each with raw request JSON, latency, tokens, and decision logic.

Can I try it without writing any code?

Yes, at jev-explained-repo.vercel.app, bringing your own TypeSafe or Vercel AI Gateway key; there is no hosted key. Locally it is npm install, npm run dev, then http://localhost:3000, where you pick a provider, paste the key, choose an example and press Run. Edit the state or switch sample states to watch the answers move.

How is my API key handled?

Keys live per provider in the browser's localStorage only, and are forwarded on each request in the x-jev-api-key header with x-jev-provider selecting the upstream. Because neither API accepts cross-origin browser calls, the app ships a small server-side proxy at src/app/api/jev/route.ts, and the server never persists a key.

What is the difference between the two providers it supports?

Only the URL, the key and the model id. TypeSafe is api.typesafe.ai/v1/systemone with model jev-latest, and the Vercel AI Gateway is ai-gateway.vercel.sh/typesafe/v1/systemone with model typesafe-ai/jev and an AI Gateway key from your Vercel team. Both speak TypeSafe's native request and response shape, so every example runs unchanged against either.

CommunityJev Demos & Experiments

Jev Explained

Interactive playground that teaches Jev by running real typed noul, choice, and score questions and showing the resulting probabilities.

See highlights

Open github.com

Stars: 34
Language: TypeScript
License: MIT
Published by: Community
Author: davila7
Last verified: Oct 2026

davila7/jev-explained

User repository on GitHub

View on GitHub

Jev Explained

Forks: 5
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Learn how TypeSafe's Jev makes typed, probabilistic decisions — by running it. Live demo: (bring your own TypeSafe or Vercel AI Gateway key). An interactive playground that shows, step by step, how Jev — TypeSafe's System One model — works.

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

Tagscommunity jevbest catalog-import

Works forJev for Classification

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Explains the three primitives with concrete spam, routing, and grading examples and their return shapes.
Four runnable patterns cover classification, structured-data calls, agent guardrails, and speculative fan-out.
Every run shows the exact request, latency, token usage, probability bars, and the code-made decision.
Supports TypeSafe and Vercel AI Gateway endpoints with editable states, samples, and raw JSON toggle.

Watch out

Bring your own TypeSafe or gateway key stored only in browser localStorage; requests go through a small local proxy route. 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

laya-server

Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.