Skip to content
Jev Hub

[ Learn ]

What is a System One model? TypeSafe's name for a decision model

A System One model is TypeSafe AI's name for a decision model such as Jev: it returns typed answers and probabilities, not text. Where the name comes from. Checked 1 November 2026.

Published By Jev Hub

[ Sponsors ]

These sponsors help keep Jev Hub updated. Want yours up here?

Back to Learn
[ TLDR ]
  • A System One model is TypeSafe AI's name for a decision model: an AI model that reads some text, which TypeSafe calls the state, and returns typed answers and probabilities instead of generated text.
  • Jev is TypeSafe's flagship and the first one.
  • The name comes from Daniel Kahneman's fast, intuitive System 1 thinking, because these models are built for quick, focused judgments.
  • TypeSafe trains them with a method it calls reinforcement learning for calibrated decisions, so the probabilities are meant to reflect real uncertainty.
  • They do not write replies, produce code or explain their reasoning.

System One models and the usual chat models

FeatureSystem One model (such as Jev)Chat and reasoning models
Trained forCalibrated decisions (TypeSafe calls this RLCD)Answers people prefer (RLHF), or verifiable reasoning (RLVR)
ReturnsTyped answers and probabilitiesGenerated text
Thinking styleFast, focused judgments, the System 1 ideaConversation, or slower step-by-step reasoning
Input today (Jev)Text only: strings, JSON objects and arrays of textOften text, images, audio and more
Who reads the outputYour software, directlyA person, or software that parses the text

The System One column is what TypeSafe's documentation says about System One models and Jev, read on 1 November 2026. The other column is the general contrast TypeSafe draws in its AI primer between RLHF, RLVR and its own method. Other makers' decision models can differ in detail.

What is a System One model?

A System One model is a class of AI models that TypeSafe AI describes as "built to make fast, structured decisions that software can use directly". It evaluates a state, the content you give it, such as a support message or the current state of your application, and returns typed answers and probabilities.

Like a chat model, it understands ordinary language. Unlike a chat model, it does not write a reply. Jev is TypeSafe's flagship model and the first System One model. Other companies have since released models that do a similar job, and the broader name for the category is a decision model.

Where does the name come from?

According to TypeSafe's documentation, the name comes from the concept Daniel Kahneman popularized in his book Thinking, Fast and Slow. System 1 thinking is fast and intuitive, and System 2 is slower and more deliberate. For these models the emphasis is on fast, focused judgments.

That is also the advice for using them. TypeSafe says to ask for a judgment a knowledgeable person could make in a second, and to break a bigger task into small questions whose answers you combine in your own code.

How is it trained differently?

TypeSafe's AI primer describes three ways of adapting a pretrained language model. Reinforcement learning from human feedback, RLHF, turned pretrained models into chatbots by training them to produce answers people prefer. Reinforcement learning with verifiable rewards, RLVR, produced reasoning models that are strong at tasks such as mathematics but slower and more expensive. TypeSafe adds a third, reinforcement learning for calibrated decisions, or RLCD.

RLCD trains the model to return decisions and probabilities, not generated text, with higher probability meant to correspond to a greater chance that the answer is right. TypeSafe argues that preference training can reward confident-sounding answers that are not reliable enough for unattended automation, which is why it picks a different target for machine use. That is the company's own position.

TRAINING METHODWHAT IT PRODUCESRLHFChat modelsanswers people preferRLVRReasoning modelsmaths, but slowerRLCDDecision modelstyped answers
Three ways of adapting a pretrained model. From TypeSafe's AI primer, read 1 November 2026. The last one is TypeSafe's own method.

What does it take in and give back?

You send a state and one or more questions. The state can be a string, a JSON object or an array of text, and TypeSafe suggests an object for most requests so each part has a name. Each question is one of three types, and each comes back as a typed answer under the id you chose.

Jev currently accepts text only. Images, audio and video are not supported, and TypeSafe says its primary training language is English, with other languages accepted at lower accuracy.

Our read

System One is a company name, and the category is a decision model, so a reader may see either one. The useful part is the contract, not the label: a typed answer with probabilities that your code can test and log.

The claims about training and calibration are TypeSafe's own, and calibration is measured across groups of predictions, not a guarantee for a single answer. Test on a labelled sample of your own data before relying on any of it.

Common questions

Sources

Written by Jev Hub from the sources below and checked against them on . Where a figure is a company's own claim, the article says so.

Related content