What Is Jev? TypeSafe's System One Model, Explained

We sent the same customer message to two AI models. It was a simple one: a customer charged twice for an order, wanting it sorted before Friday or they would dispute it with their bank.

The chatbot wrote back sixty-two polite words. It apologised, said it had located the order, and promised a refund by Friday. It had not located anything. It cannot see any order system. It simply wrote what a helpful agent would say.

The other model wrote nothing at all. It returned three numbers: this belongs to billing (100%), it is urgent (99%), and the customer's frustration sits at 1.82 on a scale of 0 to 2.

That second model was Jev. This is our plain-language guide to what it is, how it works, and where it belongs in the way we use AI.

Is Jev a chatbot?

No, and that is the whole point.

A chatbot like ChatGPT, Claude or Gemini is built to write. You ask, it answers in sentences, and a person reads them. That is wonderful when the output is for a person: an email, a summary, a plan.

It is awkward when the output is for software. A program cannot act on "I apologise for the inconvenience". It needs a value it can compare, a label it can switch on, a number it can hold against a line. So teams end up asking the chatbot to "reply only in JSON", parsing whatever comes back, and hoping it followed the instructions this time.

TypeSafe describes System One models as "built to make fast, structured decisions that software can use directly". Jev never writes a paragraph. It returns typed values and probabilities, which is exactly the shape code wants.

What is a System One model?

It is a new class of model, and Jev is the first one released. The name echoes a well-known idea from psychology: fast, instinctive judgement versus slow, deliberate reasoning. Most of the AI we use every day is built for the slow, wordy end: writing, explaining, reasoning step by step.

A System One model sits at the other end. It makes one narrow judgement at a time, quickly, and tells you how sure it is. We go deeper into the idea, and into how these models are trained differently, in System One Models Explained.

How does Jev work?

Every call to Jev has two parts, and nothing else.

The state is the evidence: a customer message, a CV, a product listing, a paragraph from a contract, or a small JSON object that bundles a few of these. The questions are what you want to know about it. Each question has a shape, an instruction written in plain words, and a description of what each possible answer means.

There is no prompt in the chatbot sense. No "you are a helpful assistant", no examples to paste, no formatting rules. The question itself carries the meaning. That also means the wording of the question matters a great deal: in our lab, changing one word in a question about the same message moved Jev's answer from 28% to 94%. We walk through that experiment, and the rest of the mechanics, in How Jev Works.

You can send many questions about the same state in one call. They are answered in parallel, and they cannot see each other's answers. In a TypeSafe test that asked 13 questions about one long document, sending them together was 12.2 times cheaper and 10 times faster than sending them one at a time, with the same answers.

What are Noul, Choice and Score?

Jev understands exactly three shapes of question. Everything you build with it is made from these.

A Noul is a yes/no question. You get back one number: the probability that the answer is yes. "Does the customer ask for a refund?" 0.99. A Choice picks one option from a list you define, up to 255 of them, and gives you a probability for every option. "Which team should handle this?" Billing. A Score places the answer on a ladder of levels you describe in words. "How severe is this bug?" 1.43, somewhere between "broken with a workaround" and "blocking".

Three shapes sounds limiting. In practice it is enough for most of the judgements a business makes all day, because bigger decisions are built by combining several small answers in code. Our guide to Noul, Choice and Score covers what each one is enough for, and the common mistakes.

What does "calibrated" mean, and why does it matter?

Think of a good weather app. When it says "70% chance of rain", it does not mean a light drizzle. It means that on days like this one, it rains about seven times out of ten. The number is honest.

TypeSafe trains Jev to make its probabilities honest in the same way. Its own documentation puts it simply: outcomes given a probability of 0.2 should happen about 20% of the time. They call the training method RLCD, reinforcement learning for calibrated decisions, and contrast it with the preference training behind chatbots, which rewards answers that sound good to people.

We tested this on 24 customer messages that we labelled by hand. When Jev said below 34%, none of the 11 messages really asked for a refund. When it said 67% or above, all 12 did. It was unsure about exactly one message, "Do you offer refunds if the workshop is cancelled? Asking for my manager", and rated it 61%. A chatbot asked the same thing said 100%, and was wrong. Twenty-four messages show a pattern, not a precise rate, but the pattern is the one that matters: when Jev is unsure, it tells you.

That is what makes the numbers usable. If 0.95 really means 95%, you can let software act alone above it, send the middle band to a person, and ignore the bottom. The number becomes a policy you can defend.

What can you build with Jev?

Anything where the same judgement about text has to be made again and again, reliably and cheaply. TypeSafe's own list is long: customer support triage, lead scoring, CV screening, content moderation, guardrails for chatbots, routing questions to the right model, reranking search results, compliance checks, insurance claims, marketplace listings and more.

The common thread is a decision that a person could make in a few seconds, made thousands of times a day, where you need to know how sure the machine was. We picked twelve jobs, with the actual questions to ask for each, in Jev Use Cases.

What can Jev not do?

TypeSafe publishes its own list of weak spots, which we respect. Jev does not count reliably, so "are there three fruits in this list?" should be split into one question per item and counted in code. It treats dates as text, so comparing two dates belongs in code too. It reads questions literally, including negations. Text written to steer it can move the answer. Irrelevant detail in the state makes it worse, not better. And it does not generate text at all: if you need a reply written, that is a chatbot's job.

It also only reads text. No images, audio or video, at least in version 1.13.

How fast is Jev, and what does it cost?

TypeSafe's homepage also claims Jev is 193.6 times faster and 444.6 times cheaper than large language models on System One tasks. We have not benchmarked that ourselves, so treat it as the maker's figure. What we can say is that a decision costing a small fraction of a cent changes what is worth automating.

Jev or ChatGPT: which one do you need?

Usually both, doing different jobs. The pattern we teach is simple: Jev decides, a chatbot writes, and your code owns the policy in between.

A support system built this way asks Jev which team, how urgent and whether a human should look. Code applies the thresholds your manager agreed. Only then, if a reply is needed, a chatbot drafts it, with the decision already made and the facts already checked. The expensive, wordy model does the one job only it can do, and every decision along the way is a number someone can inspect.

How can we try Jev without writing code?

We built Jev Lab for exactly this. It is free, it calls the real model, and it has three modes. Learn is ten short stages with a check at each one and a certificate at the end, simple enough for a 15-year-old working alone. Teach turns it into a workshop, with a live room where everyone sends messages from their phones. Build lets you paste your own messages, pick questions from a library and download a CSV plus the code to run it yourself.

If you want to see what a chatbot is doing on the other side of the comparison, Micro LLM builds a tiny language model in your browser from text your audience types. The two labs make a good pair: one shows how models write, the other how a model decides.

What does Jev mean for teams in Sri Lanka and the region?

Two things stood out in our testing. First, it read Sinhala and Tamil well. We sent the same angry refund message in English, Sinhala and Tamil, and all three came back at 99% for "asks for money back", with billing as the team every time. The Sinhala and Tamil versions did use more tokens, 588 and 643 against 514 for English, so they cost a little more to read.

Second, the price makes a different kind of automation realistic for smaller companies. Most businesses here are not short of ideas for AI. They are short of ways to trust it. A model that says how sure it is gives managers something they can sign off: act above 0.9, check the middle, ignore the rest.

This is the automation step in our 3 As of AI, and it only works once a team has learned to design the system rather than just use the tool. That is the work we do in our AI agents and automation training and, for companies that want it built with them, through our consulting.

Frequently asked questions