Jev hub: TypeSafe Jev, Noul and LitJev

Jev is TypeSafe AI’s first public System One model, announced in September 2026 after two years in stealth. You send it some state and a few typed questions. It answers with Choice, Score or Noul values plus probabilities, and your code decides what happens next. No prose, no chat.

That is a narrow job, and the posts below treat it as one. I wrote a long overview on 19 September, then a seven-part daily series from 20 to 26 September. Speed and price figures are TypeSafe’s own unless a post says otherwise. Open projects such as LitJev are independent experiments, not copies of TypeSafe’s model.

Start here

• LitJev and TypeSafe Jev: System One decisions for software: the overview. What Jev returns, where the “no hallucination” claim stops, launch prices and the demos.
• What Noul means in AI: Jev’s yes/no probability data type: start here if you searched for Noul.
• Jev ecosystem: gateways, open replicas, and the limits: access through OpenRouter, Vercel and Cloudflare, plus LitJev and the other open projects.

Reading order: the seven-part series

The series ran daily from 20 to 26 September 2026. Later parts refer back to earlier ones, so the original order works best.

1. Day 1: Choice, Score and Noul (20 Sep). The three question types, field paths and one call that uses all three.
2. Day 2: Confidence and RLCD (21 Sep). How confidence comes from the probability distribution, and how to scale thresholds to risk.
3. Day 3: Speed and cost (22 Sep). The parallel sampler, the 70 to 500 ms figure and the $0.042 per million input tokens price.
4. Day 4: Agent guardrails (23 Sep). Why a schema-valid tool call can still be unsafe, with jev-guard and Canny as examples.
5. Day 5: Context compaction (24 Sep). Pruning agent history with Noul questions and keeping what survives word for word.
6. Day 6: Real-time demos (25 Sep). Doom, Wikiracing and browser loops, all running on text state rather than pixels.
7. Day 7: Ecosystem and limits (26 Sep). Gateways, open replicas and the jobs that should stay with an LLM.

Want the context first? Read the overview from 19 September before day 1.

Concepts and primitives

Jev takes state plus named questions and evaluates the questions in parallel. The name comes from Daniel Kahneman’s System 1 and System 2: fast, narrow judgments go to Jev, while open-ended reasoning stays with language models.

• TypeSafe Jev overview: the state-plus-questions request, the System One framing and TypeSafe’s jev-1.13 jaggedness notes.
• Jev primitives in working code: when to pick Choice, Score or Noul, and how to write a question the model can answer.
• Jev confidence and RLCD: confidence is derived from the shape of a Choice or Score distribution. RLCD is TypeSafe’s name for its training method.

Noul data type

Noul returns the probability that a yes/no proposition is true, as a number from 0 to 1. It has no separate confidence field. People search for “Noul data type”, but Noul is a question type in TypeSafe’s API, not a built-in type like a Python bool.

• What Noul means in AI: the full explanation, with Python code and advice on thresholds.
• Why Noul has no confidence field: covered in the confidence post.
• Noul in practice: context compaction: fast-jev-compaction asks two Noul questions per tool call.
• Noul as a refund gate: OpenRouter’s example in the guardrails post uses three Noul questions with 0.9 and 0.1 thresholds.

LitJev and open replicas

LitJev wraps off-the-shelf Qwen checkpoints with no extra training and scores typed choices directly, without generating answer text. Its README calls it a hypothesis-based reproduction and says the probabilities are not calibrated by default. It is not TypeSafe’s Jev. Neither are openjev, Kev or NanoJev.

• LitJev: the zero-training Jev reproduction on Qwen (9 Oct). Dedicated walkthrough of the Apache-2.0 local /v1/systemone reproduction.
• Jev ecosystem: open replicas: LitJev, openjev, Kev and NanoJev side by side, with licenses and model sizes.
• The community map: the awesome-jev list had passed 100 projects by 19 September.

Agents and guardrails

• Jev agent guardrails: schema-valid is not the same as safe: deterministic checks first, then probabilities, then allow, ask or deny.
• Jev context compaction: what the model may rank and what host code must keep.
• Browser agent loops: serialized DOM and accessibility-tree state, from the real-time post.

Speed and cost

• Jev speed and cost: 70 to 500 ms end to end in TypeSafe’s listing, free output tokens, and why the 193.6x and 444.6x figures are ceiling claims.
• Doom at about $7 an hour: what a 10 Hz loop costs and how rate limits shape the design.
• Batching questions: the cookbook run where one batched call came out 12.2 times cheaper than 13 single calls.

Ecosystem and providers

• Clef vs Jev: Cloudflare’s open-weight decision models (10 Oct). Cloudflare’s own Jev-compatible models compared on price, the latency-method caveat, small independent tests and where each fits.
• Three gateways: OpenRouter’s Decisions API, Vercel AI Gateway, Cloudflare Workers AI and TanStack AI’s decide() layer.
• Provider IDs and version pinning: typesafe/jev-1.13 on OpenRouter, typesafe-ai/jev on Vercel, typesafe/jev on Cloudflare.

Real-time demos

• Jev real-time: Doom, Wikiracing and browser loops: text state, the 255-option Choice limit and the score-then-choose pattern.

Jev in the wider AI news

• AI weekly, 14 to 20 September 2026: the Sunday bulletin’s side note on decisions that should never become prose.
• OpenAI DevDay 2026: OpenAI’s Decisions API, which Simon Willison read as a response to Jev.

FAQ

What is Noul in Jev?

One of Jev’s three question types. You give it a yes/no proposition and it returns the probability of yes, from 0 to 1. A value near 0.5 means the model is split, not that the answer is “medium”. Full explanation.

Is Noul a data type?

Not in the programming-language sense. It is a typed question and answer in TypeSafe’s API. The answer is a number, but Noul is not a built-in scalar like bool or float, and “Noul programming” is not a separate language. Vercel’s AI SDK calls the same shape Boolean.

What is LitJev?

An open, Apache-2.0 project that produces Jev-style typed decisions from off-the-shelf Qwen checkpoints without training. Its own README says the probabilities are not calibrated by default. Dedicated LitJev post; ecosystem map of the other replicas.

What does System One mean for Jev?

TypeSafe borrows Kahneman’s split between fast judgment and slow deliberation. Jev handles narrow, quick decisions inside software. Language models keep the writing, planning and open-ended reasoning. The overview covers that boundary.

How fast and how cheap is Jev?

At launch TypeSafe listed roughly 70 to 500 ms end to end and $0.042 per million input tokens, with output free. Those are company figures, so test on your own traffic. Speed and cost.

What is RLCD?

Reinforcement Learning for Calibrated Decisions, TypeSafe’s name for how it trains Jev. I have not found an independent reproduction of the full method. Confidence and RLCD.

Beyond Jev

Daily AI digests and deep dives live in the Artificial Intelligence category, and security stories sit under Cybersecurity. New Jev posts get added to this page the day they go live.