What is Jev?
Jev is TypeSafe AI's first public System One model: it takes state plus typed questions and returns bounded decisions with probabilities instead of free-form prose.
State + typed question → decision + probability
Illustrative example flow — this is not a live Jev response.
Three ways to ask Jev
Pick one label from a closed set.
Routing · classification · tool selection
Estimate the probability of a yes/no proposition.
Verification · eligibility · completion
Place the state on an ordered rubric.
Urgency · risk · quality · priority
Probabilities make close calls visible
A high-margin result can be routed automatically if your policy allows it.
A close distribution is a signal to clarify, fall back, or ask for review.
Try the idea before reading more
The Jev Playground lets you edit the state, switch between Choice, Noul, and Score, and inspect the returned decision, confidence, and probability distribution.
Jev in one minute
A normal generative LLM can produce almost any string. Jev gives up that open-ended output space. Your application defines the shape of the decision in advance, and Jev fills that shape with a typed answer and uncertainty information.
import { choice, TypeSafeClient } from "@typesafe-ai/sdk";
const client = new TypeSafeClient();
const result = await client.systemOne({
state: { message: "I was charged twice. Please fix this ASAP." },
questions: {
category: choice("What is this ticket about?", {
billing: null, technical: null, other: null,
}),
},
});The three question primitives
Choice
Choice selects one label from a named set. Use it for categories, routes, tool selection, or any other mutually exclusive branch.
Score
Score evaluates state against an ordered rubric. The official JavaScript SDK represents the rubric as a list with at least two score levels, indexed from zero.
Noul
Noul is a yes/no question. The returned noul value is the probability of the yes outcome. It is useful for checks such as “does the evidence establish completion?” or “is this message a qualified lead?”
Where Jev fits well
- Classifying support, sales, moderation, or document state.
- Routing a request across a bounded workflow graph.
- Scoring against an explicit severity or priority rubric.
- Filtering large volumes before sending a smaller set to an LLM.
- Verifying claims or agent actions against supplied evidence.
Where Jev is not the right tool
- Writing an email, article, program, or free-form explanation.
- Creative generation with an open output space.
- Tasks where the valid answers cannot be defined or decomposed.
- Final authorization for high-impact actions without deterministic policy checks around the model.
Jev and LLMs can be used together
A common architecture is to let Jev classify, route, or filter first, then call a generative model only when you actually need generated language. That keeps the decision layer bounded while preserving the flexibility of GPT, Claude, or another LLM where it matters.
Primary sources
TypeSafe's launch post describes Jev as “structured state in, typed probabilistic decisions out.” The official JavaScript SDK exposeschoice, score, noul, andsystemOne.
Sources last reviewed: 2026-09-21.
Next
Try the Startup Idea Validator, inspect the raw decisions in the Playground, continue with the getting-started guide, estimate usage in the cost calculator, or browse practical templates.