Jev and the rise of judgement models: what can they really do?

In mid-September, TypeSafe AI launched Jev in early access, the first model in a class the company calls “System One Models”. The demos spread quickly: an AI model that returns decisions instead of text, in 70 to 500 milliseconds, at a fraction of the cost of a large language model. You may well have seen something about it or had a colleague forward it to you, complete with bold claims such as “0% hallucinations”.

The hype is not entirely unjustified, because there is a real idea behind it: a judgement model (in Dutch: beslissingsmodel) chooses from predefined answers and returns a confidence score, instead of writing text. But the claims deserve a sober reading, and we will examine that zero per cent more closely below.

Writing versus choosing

Large language models such as ChatGPT and Claude are built to produce text: answers, summaries, emails, code. That is impressive, but also slow and relatively expensive, and the answer comes as free text that your software then has to interpret.

A judgement model does something different. You give it a fixed set of possible answers in advance, such as a list of categories, a scale from 1 to 10, or a simple yes/no. The model chooses from those options and returns a confidence score: how certain it is of that choice. You get nothing more back, and that is exactly the point. The name System One refers to Daniel Kahneman’s distinction between fast, intuitive thinking (System 1) and slow, deliberate thinking (System 2): this model is the fast system, for decisions that do not need an explanation.

Many steps in our work processes can be expressed as that kind of choice. Which category does this complaint fall into? Which department should this email go to? Does this invoice line belong to purchasing or marketing?

For steps like these, a full language model is more than you need.

What Jev can actually do

According to TypeSafe’s own documentation, Jev is built for tasks such as classification, routing, scoring, identifying fields in text and checking the output of other AI models. The company claims response times of 70 to 500 milliseconds, at costs that work out to a fraction of a cent per decision. Those are vendor figures, but if they are even close to accurate, the difference compared with large language models is interesting. In practice, you can use a model like this where a language model is too slow or too expensive: in the middle of a work process, for every incoming document or every customer enquiry.

The most interesting part is the confidence score. It enables a way of working we already use in our AI agents: decisions the model is confident about are processed automatically, and anything below a threshold goes to a person. You automate the bulk of the work while uncertain cases land on someone’s desk. For most SME processes, that is a much healthier setup than all-or-nothing automation.

What it cannot do, and what the claims really mean

The other side matters just as much. Jev does not write text, so it cannot formulate an answer, give an explanation or ask a follow-up question. It only reads text, not images or scans, and the size of the list of options is limited. The model is also trained primarily on English, so for Dutch texts, testing on your own material is especially important.

Then there is the best-known claim: “0% hallucinations”. As MarkTechPost rightly notes, this only means that the answer is guaranteed to come from your predefined list. The model therefore cannot invent a category that does not exist. But it can still choose the wrong category. The guarantee concerns the form of the answer; the claim says nothing about whether the choice is correct.

The same trade press therefore advises against blindly accepting vendor benchmarks and recommends testing on your own data. We fully endorse that advice.

Hosted or self-hosted

Jev is a hosted API with a waiting list: your data goes to TypeSafe’s servers in the United States, according to its own privacy policy. OpenAI has now also entered this market: at Dev Day (29 September 2026), it announced the Decisions API, built on GPT-6 Luna, which can assess images as well as text but is currently only available in closed preview and does not yet publish a confidence score. If you prefer not to send anything outside your environment, open-source alternatives exist in the same category, such as Laya. It runs locally, so your data stays in your environment, but it must first be fine-tuned on your own domain. The right option depends on the sensitivity of your data and how much evaluation work you can take on yourself.

What this means for your organisation

The lesson from Jev goes beyond the product itself: not every AI step in a process calls for a large language model. Many steps are choices, and there is now a category of models for those choices that is fast, inexpensive and indicates when a choice is uncertain. Combining both types of models, with a person overseeing uncertain cases, is how we currently see most working automations taking shape.

We received early access to Jev ourselves and have used it for two internal agents: we tested it as a decision layer for our accounting agent Penny and built it into our business agent Twine as a router. In this look behind the scenes, you can read what we built, where Jev stops and where a language model or person takes over.

Want to know which decisions in your own processes could suit this approach? Our AI Readiness Scan identifies where AI can make a difference in your organisation, including where a fast decision layer like this fits.

A note on how this was made: our content is AI-assisted. We use AI for research and drafting a first concept. A human reviews and edits the text and takes responsibility for everything we publish. Read how we make our content.