Jev
TypeSafe AI’s first “System One” model, launched in early access in September 2026. It answers typed questions about a piece of state and returns structured answers with probabilities. It does not write text.
Glossary · 14 terms
Jev’s vocabulary, in plain English
Each definition links to where TypeSafe defines the term, so you can check our wording against theirs.
System One model
TypeSafe’s name for models built for fast, structured decisions rather than long reasoning or writing. The name borrows Daniel Kahneman’s “System 1”, fast thinking.
State
The input Jev judges: a ticket, an email, a transcript, a JSON object or an array of text. Text only; no images, audio or video.
Question
One typed thing you want decided about the state. You send many at once, each with an ID you choose, and they are answered in parallel and in isolation.
Choice
A question whose answer is one of a set of options with no order between them, such as which team gets a ticket. Returns the pick, a probability for every option, and a confidence.
Score
A question whose answer sits on a scale you describe level by level, such as how frustrated a customer sounds. Returns a position on the scale (it can fall between levels), the probabilities and a confidence.
Noul
TypeSafe’s yes/no question. Returns one number from 0 to 1: the probability the answer is yes. Vercel’s gateway calls this type “boolean”.
Confidence
A 0-to-1 summary of how concentrated a Choice or Score answer’s probabilities are: 1 when all the probability sits on one answer, 0 when it is spread evenly. TypeSafe does not claim it equals real-world accuracy.
Confidence-gated routing
Letting code act on confident answers and sending unsure ones to a person or a stronger model.
Composite scoring
Asking several narrow questions at once and combining the answers in code, for example with weights, instead of asking one broad question.
Speculative fan-out
Asking questions you may or may not need in the same request, because extra questions run in parallel and add little time.
RLCD
Reinforcement Learning for Calibrated Decisions, TypeSafe’s name for how Jev is trained. It aims for honest probabilities rather than fluent text.
Jaggedness
TypeSafe’s page listing where the current Jev version is weak: counting, raw numbers, dates, long noisy state, double negatives, option order and adversarial input.
jev-latest
The model alias most code uses. As of October 2026 it points to jev-1.13.0, priced at $0.042 per million input tokens with free output.