System One models

System One is a class of models built for fast, structured decisions. Jev is the first, and its output can be consumed by software directly.

Definition

System One models are a class of AI models built to make fast, structured decisions that software can use directly. Jev is TypeSafe’s flagship model and the first System One model.

The problem they solve: conventional language models emit free-form text, while software needs values of a known type. System One internalises that conversion — the question declares the output type, and the model answers within it.

Division of labour with System Two

The naming borrows from dual-process theory in cognitive science, and the meaning is direct:

	System One	System Two
Character	Fast, intuitive, focused	Slow, deliberate, multi-step
Typical tasks	Judgement, classification, scoring, verification	Complex reasoning, long-horizon planning
Latency	Low and predictable	Higher, grows with thinking length
Output	Typed and constrained	Free-form text
Cost	Low	High

They are not substitutes. The typical production pattern is for System One to carry the overwhelming majority of high-frequency judgements, escalating to System Two or a human only when deep reasoning is genuinely needed.

A worked example: refund requests

The flow in the official docs illustrates how the two layers cooperate:

Build a state — pack the customer’s message, the relevant transactions, and the refund policy into one state.
Ask in parallel — ask three independent questions at once: whether a refund was requested, whether the evidence indicates a duplicate charge, and whether the policy supports a refund.
Combine in code — combine the three answers with deterministic business checks, then route to action or review.

Note the independence of the questions in step 2: they do not influence each other, which is the precondition for packing them into a single request.

Why typed output is the crux

Because System One models return typed, constrained outputs rather than free-form text, your code can inspect and combine the answers into predictable workflows.

Compare the integration complexity:

# Conventional: parse, validate, handle format drift
raw = llm.complete("Is this ticket about billing? Answer yes or no.")
is_billing = raw.strip().lower().startswith("y")  # brittle, many edge cases

# System One: the value is already typed
response = client.system_one(
    state=ticket,
    questions={"billing": Noul(instructions="Is this ticket about billing?")},
)
is_billing = response.nouls["billing"].noul  # float, 0..1

The second returns a type within a known domain — a Choice is always one of your options, a Score lies within your levels, a Noul is always a float from 0 to 1.

Confidence: letting the model say “I don’t know”

System One answers also carry confidence, so you can decide when to act and when to escalate to a person or a reasoning model. This is the foundation of trustworthy systems — if a system cannot express honest uncertainty, it cannot be trusted.

How to call it

Through a client SDK or the HTTP API:

POST https://api.typesafe.ai/v1/systemone

The model field in the request selects which model handles the call. The default alias is jev-latest.