What is a System One model?
A System One model is TypeSafe’s name for a class of models built to make fast, structured decisions that software can use directly. The official concept page says a System One model evaluates a state and returns typed answers and probabilities. Jev is the flagship and the first public model in that class.
This page is for builders who already heard “System One” and need the class definition before they wire What is Jev?. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. Confirm the concept page on docs.typesafe.ai. We do not sell or issue API keys.
Definition
Like an LLM, a System One model understands natural-language input. Unlike an LLM, it does not emit free text. You define the possible answers through primitives: Choice, Score, and Noul. Questions in one request are evaluated independently and in parallel against the same state.
The useful contract is a function call: unstructured state in, typed probabilistic decisions out. No text generation. English is the primary language. Text only. Current public instance: Jev, versioned id jev-1.13.0.
Why the name (Kahneman, not marketing fog)
TypeSafe says the name comes from Daniel Kahneman’s Thinking, Fast and Slow. System 1 is fast and intuitive; System 2 is slower and more deliberate. The emphasis here is on fast, focused judgments — the kind of call a knowledgeable person could make in a few seconds given the right context.
The launch post (15 September 2026) notes that “System 1 thinking” has also implied error-prone. TypeSafe writes that it believes System One models can be made more reliable than the alternatives, and defers the longer argument. Treat that reliability claim as vendor belief, not a bench this site ran.
Jev is the first public instance of the class. “System One” is not a second product you buy. It is the category name.
Mental model: code owns System 2
| Layer | Owner | Example |
|---|---|---|
| Snap judgment | System One (Jev) | Was a refund requested? |
| Slow checks | Your code | Entitlement, velocity, ledger math |
| Writing | An LLM, if needed | The customer email |
Do not ask Jev to be System 2. Keep arithmetic in code — jev-1.13 is weak at numeric precision, extra indirection, and overly literal wording (jaggedness). Schema-safe ≠ correct: a typed answer can still be the wrong call.
How TypeSafe contrasts it with LLMs
From the System One page and the launch comparison:
- Training target. System One models are trained for calibrated decisions (RLCD). Probabilities are optimized against outcomes to reflect uncertainty. Calibration is measured across groups of predictions; it does not guarantee that any single answer is correct.
- Outputs. Possible outputs and structure are defined in advance. The model returns values inside that structure.
- Sampling. TypeSafe describes parallel evaluation of questions rather than token-by-token generation.
- Role in a product. Code owns control flow. The model supplies snap judgments.
Vendor-published envelope for Jev (confirm on the models page): $0.042 per million input tokens, output free; 70ms–500ms; “40x–200x”; 64k / 32k context. Label those figures.
A refund-shaped workflow
The official page walks a refund request:
- Build a state with the customer message, transactions, and policy.
- Ask independent questions together: was a refund requested, does evidence show a duplicate charge, does policy support a refund.
- Combine answers with deterministic checks, then route for action or review.
That pattern — decompose, ask in parallel, compose in code — is the System One design loop. See atomic questions and how to build.
See also
FAQ
Is System One a separate API from Jev?
No. Public docs evaluate via POST /v1/systemone. Jev is the model you name. System One is the class.
Why Kahneman if System 1 is “error-prone”?
TypeSafe borrowed the fast judgment half of the metaphor and argues the models can be made more reliable than alternatives. That argument is theirs. Measure your traffic.
Does calibration make one refund decision correct?
No. Calibration is about groups of probabilities. Schema-safe ≠ correct. Your code still needs floors and review.
Can a System One model write the denial email?
Not Jev. Generation is out of scope. Compose an LLM after the decision if you need prose.
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.