Jev and "System One" models
TypeSafe's Jev returns typed, probabilistic decisions instead of text. A visual walkthrough: why it exists, how it differs from an LLM, what people are building, and what's still unproven.
TL;DR: TypeSafe AI launched Jev on September 15, 2026. It is not a chatbot and not a coding model — it never writes text. You hand it some state plus a list of typed questions, and it returns one typed answer per question with a probability attached. The bet: language generation is the wrong interface between a model and the software that has to act on its output.
1. Start with what code can't do
Ordinary code branches on things it can compute. It falls apart the moment the condition is a judgment.
So we hand the judgment to an LLM. And that's where things get messy.
2. Six things that go wrong
"billing". You get a paragraph. So you beg for JSON, then parse, validate, retry.We're using a text-writing machine for what is really a multiple-choice question. The interface is language — and language was designed for humans, not for if statements.
It's hiring a novelist to fill in a bubble sheet.
3. So how would you fix it?
Follow the questions down, one at a time:
What does software actually need?
Data in, an answer it can use directly out. Nothing to parse.
If the answers come from a fixed menu, do we need words at all?
No. Just score every option on the menu.
If nothing is being written, must it go token by token?
No. Answer every question in parallel, in a single pass.
And can it still go off-menu?
No. It can only return one of your options — so invented categories and type errors become structurally impossible.
LLM as decision-maker the long way round
Jev the short way
4. Watch the difference
The same decision, two ways. One writes it out; one just picks.
5. What Jev actually is
TypeSafe's framing: unstructured state in, typed probabilistic decisions out — a frontier-intelligence function call. One writer put it better: a smart if statement. Three answer shapes:
- choice — pick one of 1–255 options you defined, with full probabilities
- score — a position on a rubric of 2–10 ordered levels
- boolean — the estimated probability that something is true
const answers = await jev({ state: supportMessage, questions: { topic: choice(["billing", "bug", "refund", "other"]), anger: score(5), needsHuman: boolean(), }, }); // answers.topic -> "billing", p = 0.93 // no parsing, no "please reply in JSON"
Your code acts when confidence is high and escalates when it isn't. Try it on a few messages:
Illustrative only — these outputs are hardcoded to show the shape of a response, not a live API call.
6. Where the name comes from
Kahneman's Thinking, Fast and Slow. System 1 is fast instinct — catching a ball. System 2 is slow deliberation — solving a puzzle. LLMs with chain-of-thought and multi-second reasoning traces sit firmly in System 2 territory.
TypeSafe's argument is that the automation gap exists because everyone built System 2 models, while most decisions inside software want a System 1 answer.
7. How it differs from an LLM
| Typical LLM | Jev | |
|---|---|---|
| Output | free-form text | typed answers from your menu |
| Generation | one token at a time | all questions in parallel, one pass |
| Confidence | vague, in words | calibrated probabilities |
| Chat or write code? | yes | no, by design |
| Best at | open-ended thinking, writing | fast repeated decisions, known answers |
Against older AI — a sentiment classifier trained for one fixed task — the difference is flexibility: you define new questions at call time instead of training a new model per question. The training method is RLCD (Reinforcement Learning for Calibrated Decisions), aimed at making those probabilities honest.
8. The numbers
(output free)
9. What people built in week one
The pattern worth noticing: in the router, the browser agent and the safety monitor, Jev doesn't replace the LLM. It wraps around it as a reflex layer — deciding who answers, what to click, whether an action is safe — and hands off to a big model only when needed.
A model that can't write paragraphs turns out to be useful for the small decisions between the paragraphs.
10. Caveats
- "Cannot hallucinate" is narrower than it sounds. It can't emit an invalid answer or a type error. It can still pick the wrong option off the menu. Valid ≠ correct.
- All benchmarks are TypeSafe's own. No large-scale independent reproduction has surfaced yet.
- Measure cost per workflow, fallback models and human review included — not per cheap API call.
- It is days old. Everything here is demos and launch-week enthusiasm, not production track record.
The idea underneath
Not every intelligent part of a system needs to talk. Some tasks need a paragraph; others need a bounded decision, usable uncertainty, and an explicit account of what happens next.
Whether RLCD delivers on that is unproven. But the question is a good one: how much more dependable could AI get if we stopped requiring every intelligent component to speak?
✦