Jev: the AI judgment model
that generates no text
TypeSafe AI's flagship System One model — send a state and a batch of typed questions, get back structured answers with calibrated probabilities. Code branches directly. Millisecond latency. A fraction of LLM cost. This is our independent guide to it.
Start here
Independent analysis and field guides — not a copy of the official docs. Every piece is written to answer a question a builder actually has.
Why Jev generates no text
The 'decision layer' thesis: why most production AI is judgment, not generation — and what RLCD actually changes.
Read →Pricing & cost math
$0.042/Mtok, output free. A transparent, assumption-driven cost-per-million-judgments comparison vs LLM + JSON mode.
Read →Choice / Score / Noul field guide
The three question types, when to use each, and the design rules that separate working calls from flaky ones.
Read →Patterns, with working code
Speculative fan-out, confidence-gated routing (full Python), composite scoring, intent routing — the four recipes that cover most builds.
Read →Known limitations & workarounds
The official jaggedness list, each with a concrete production mitigation — including the CJK accuracy note.
Read →The ecosystem, mapped
Open replicas, tooling, and demos from the first month — and what the shape of it tells us about the category.
Read →Latest from the news timeline
Dated, concrete updates on Jev and the System One space.
What is Jev, in one paragraph?
Jev is the first System One model: it makes fast, structured judgments instead of generating text. You send a state plus typed questions (Choice / Score / Noul), it returns answers with calibrated probabilities in a single forward pass, and your code branches on them. Trained with RLCD (RL for Calibrated Decisions), it targets the decision layer of the AI stack — routing, scoring, gating — where LLMs are slow, expensive, and uncalibrated. Read the full independent take in What is Jev?