From Jev-curious to Jev Pro

What Jev is

Jev is TypeSafe AI’s flagship System One model: a non-generative decision layer. You send unstructured state and a map of typed questions; you get back structured answers your software can branch on — not a chat reply, not generated code, not an explanation of its reasoning. The three primitives are Choice (which option), Score (which level on a rubric you define), and Noul (P(yes) in 0–1). Choice and Score also return confidence. TypeSafe’s contract is schema-matching: a Choice cannot invent an option you did not list. It can still pick the wrong one.

TypeSafe names the class after Kahneman’s System 1: fast, focused judgments a knowledgeable person could make in a few seconds. They train Jev with RLCD — reinforcement learning for calibrated decisions — not for preferred essays. Calibration is a property of groups of predictions; your thresholds and escalation policy still belong to you. Read What is Jev, System One, Choice / Score / Noul, and RLCD when you want the longer briefs — not the whole cluster on this page.

When to use it — and when not

The one-line test: could a sharp teammate answer this in a few seconds from this text, and can your code act on a label? If you need a paragraph, a proof, a picture, or a regulated attestation, it is out of scope. Decide with Jev; draft with an LLM. The longer boundary note is When to use Jev.

How Pros think about it

Five moves, in order. Each line is one job — not a cluster dump.

1. Explain

   Get the decision-model frame first: Jev is a function call with a closed answer space, not a chat you prompt into JSON. What is Jev.

2. Integrate

   One POST to /v1/systemone, pin a versioned model, and take keys only from TypeSafe — this site does not issue them. First API call.

3. Compose

   Ask atomic questions, combine answers in your code, and keep generation in an LLM (decide, then draft). Atomic questions.

4. Compare

   Put Jev next to JSON-mode, classifiers, and prompt-only stacks on purpose: schema-safe is not the same as correct. Jev vs prompt-only.

5. Operate

   Pin the model, set floors, escalate the uncertain, score policy actions on a harness, and fail closed. Operate practices.

First working shape

One evaluation is POST https://api.typesafe.ai/v1/systemone with a bearer key from TypeSafe’s console — never from this site — plus model, state, and a questions map. Confirm the live contract on docs.typesafe.ai/api.

POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer $TYPESAFE_API_KEY
Content-Type: application/json

{
  "model": "jev-latest",
  "state": "<ticket or packet>",
  "questions": {
    "urgency": {
      "type": "noul",
      "instructions": "Does this message express urgency?"
    }
  }
}

A successful body echoes a resolved model, an answers map keyed like your questions (here answers.urgency.noul in 0–1), and usage. Mix Choice, Score, and Noul in the same request; you do not need three HTTP calls. Walkthrough: your first API call. Official quick start: docs.typesafe.ai/introduction/quickstart.

Pro checklist (teaser)

Self-check before you treat a call as production. Full lists live on the operate pages.

• Pin a versioned model (today’s public id: jev-1.13.0) once floors exist; log the resolved model.
• Treat calibration as a group property; set your own auto / review / abstain floors — confidence and escalation.
• Escalate low-confidence or high-stakes answers. Do not auto-act on a lone 0.8.
• Score policy actions (act / review / abstain) on a frozen harness, not only argmax — evaluation guide.
• Keep arithmetic, dates, and counts in code. That jagged edge is documented for jev-1.13.
• Persist a decision trace; fail closed on 401 / 422 / 429 / 529.
• Do not paste vendor speed or price figures into a dashboard as if you measured them. Named failure modes: anti-patterns · implementation checklist.

Stay current

Latest briefing: Oct 4, 2026 · All briefings · Newsletter. Independent notes — not TypeSafe product news.

Top Repos

A peek at public GitHub stars — Jev-Based (TypeSafe Jev as the core), Jev Integrations (larger products that call Jev), and OSS alternatives (System One–class replicas that are not TypeSafe Jev). Rank is stars, not a quality award. Full tables live on Top Repos.

See all three tracks and every ranked repo — 1d / 7d / 30d live on the hub.

Browse the guide

Substance first; these are exits into the guide — not another directory of every cluster.

• ExplainersMental models, System One, primitives, myths.
• APIEndpoint, request shape, SDKs, first working call.
• Use casesRouting, triage, guardrails, and other workflows.
• ComparisonsChoose against JSON-mode, classifiers, and jagged edges.
• Top ReposJev-Based, Jev Integrations, and OSS alternatives (not Jev).
• BlogDaily JEV Briefings — same note as the newsletter.