Jev Answers in Probabilities. The Question Is Still Yours.

For about a week my feeds have been full of a model that does not write. TypeSafe AI calls Jev a System One model. You send it a state, any text or JSON, and typed questions. A Choice picks one of up to 255 options you defined, a Score rates the state on two to ten levels you described, and a Noul returns the probability that a statement is true. What comes back is the selected answer, a probability for every answer you allowed and, except on a Noul, a confidence value. No prose. TypeSafe prices it at $0.042 per million input tokens with output free, and quotes 200 times faster and 400 times cheaper than the chat models people use for the same jobs.

TypeSafe paused signups before I had a key, and the essays here are supposed to measure rather than testify. So this one weighs the writeups of people who did try it, then runs the pilot I would want before believing any multiple, on my own essays, with the open model that stood in for Jev.

The interesting part is the shape, not the speed

Strip the multiples away and what is left is a contract. A conversational model returns text your code must parse, validate and second-guess. Jev returns something a program can branch on directly, and refuses to return anything else. The model supplies a judgment. The application owns what happens next.

In the agent experiment the workflow's first step was a Jev question before Jev existed: the model picked a route from a closed enum, the arithmetic was code, and so a correct classification could not produce a wrong number. Jev makes that pattern the only thing the model can do. It is the first hyped model whose selling point is what it will not do.

It fits repeated judgments whose answer space you can name: which team owns this ticket, does this proposed action look destructive, did the task complete. It does not fit anything exact, like arithmetic or dates, or the customer-facing reply. And it does not make your agent safe. It makes one kind of decision cheap enough to ask about everything.

The number everyone quotes is the wrong one

Two hundred times faster and four hundred times cheaper are the vendor's figures for selected comparisons. One writeup ran four judgments over a thousand app reviews for two cents, against sixteen for a fast Gemini model, then noted that it had not compared accuracy at all. A cheap decision endpoint is not a cheap workflow. What you price is the completed task, including the fallback model, the human review and the tasks that went down the wrong branch.

The number that matters sits underneath the multiples. When I built Smithproof's control plane, the lesson I wrote down was that the model you can afford to run inline is the model that governs. Jev bids for that seat. A single decision over a thousand tokens costs four thousandths of a cent. A generative judge is something you run on a sample and read. A decision model is something you run on every trace and every proposed tool call, and branch on. Screen everything cheaply, and spend the expensive model on the failures the cheap one flagged.

Confidence is not correctness

The probability is the part of Jev that looks most like recourse and is least like it. The confidence field is arithmetic on the spread, and no outcome enters it: Archer Hume, who worked out Jev's architecture from about ten thousand API calls, found the formula by reading TypeSafe's own adapter. TypeSafe trained for calibration, and as I write has published no calibration curve and no paper. And the vendor's own accuracy figure is agreement with the averaged answers of two frontier chat models, 67.8% overall and 61.8% on invoices. Agreement is not correctness.

A spam test written up on Arize's blog shows what the probability is for: zero-shot Jev was right on 98.3% of the emails, the 4.6% whose probability fell between 0.3 and 0.7 went to a person, and the rest came back 99.5% right. The probability did not make the model right. It bought a review lane and said how wide to make it.

Take the example that circulates: a support ticket with billing at 0.52, technical at 0.46, sales at 0.02. Billing wins with a confidence of 0.28. It is a coin toss wearing a label, and a system that branches on the winner has automated a coin toss. Jev gives you the number and withholds the reasons, so the checking has to happen outside the model: measure correctness by probability band on your own labeled examples, set each threshold by what an error costs, pin the model version, and re-test whenever the rubric, the model or the input distribution moves.

A schema-valid mistake is still a mistake

The marketing says Jev cannot hallucinate, which is true in the narrow sense that it cannot answer outside the space you defined. A valid answer can still be the wrong answer. Jev removes the malformed output, which was always the cheap failure to catch. It keeps the confident wrong branch, which was always the expensive one, and takes away what a generative model gave you for diagnosing it: an explanation you could read.

So the recourse has to be built. Hard rules go first. The author of jev-gate, a permission gate for coding agents, ran roughly three hundred calls against it: a blunt injection got none of thirty dangerous commands through, and a line claiming the repository owner had already approved the command got three. Three of thirty makes Jev a filter in front of a boundary, never the boundary, which is why the hard rules sit in front of it. The uncertain band goes to a person or a model that can talk; a Noul cannot abstain, so the abstention is your threshold. And every question gets an exit option, none of these or needs review, before you tune rather than after. The pilot below shows why.

Then the question I asked ten vendors last month: if the 0.52 is wrong and it costs me something, who pays? I have not read TypeSafe's terms yet, and I would be surprised if they read differently from the ten I did: warranted for nothing, verify everything, a cap near the price of a paperback. A probability does not change that, so the person who set the threshold is the person who answers.

The pilot, run on my own index

Jev I could not run, so the pilot ran on the stand-in. Within days of the launch the number one trending model on Hugging Face was Laya, from Convai Innovations. It has the same shape: a state goes in, the same three question types, and the same fields come back. It is Apache 2.0, and at 421 million parameters it runs on this laptop. The numbers below are Laya's, not Jev's.

The workload is the one labeled set this blog owns, its index: 33 essays, each filed by hand under one of six subjects and one to four of the four addresses. So "agreed" below means agreed with my filing, and disagreement is not automatically the model's error. Before any number I wrote down what an error costs, a reader who presses a subject chip and lands on the wrong piece, and then the bar: beat the commonest label on every question and keep the answers when the options are reversed. Each essay got six questions: a Choice for the subject, a Choice for the address it argues most directly, and four Nouls, one per address. No exit option, because the index has none. Each went out in three wordings: the base, the options reversed, and a paraphrase. Laya ships three checkpoints, English, multilingual and one fine-tuned on typed decisions, and each ran as shipped and again with its context widened until whole essays fit. As shipped, the English checkpoint reads 512 tokens; once the subject question takes its share, 316 are left for the essay, and the median essay here runs to 2,658 tokens. Nothing in the answer says that most of the essay went unread. A hundred and sixty lines of Python serve Laya in Jev's wire format, so the same harness sends the same requests, and all six configurations ran in under a quarter of an hour of GPU.

The best result is right half the time, and cannot say which half. The checkpoint fine-tuned on typed decisions, reading whole essays, agreed with my filing on the subject for 17 of 33 essays; always answering the commonest subject would get 8. On the Nouls its Brier score was 0.257 in the base wording and 0.308 paraphrased, against 0.231 for always answering the base rate, which is worse than a constant. And it knew. Of its 132 yes/no probabilities, 131 fell between 0.3 and 0.7, and every subject came back with a confidence below 0.3. Another configuration was confident instead: the multilingual checkpoint reading whole essays put 119 of 132 above 0.7 and was right on 37%, below the 64% that always answering no would score. That is the confident wrong branch, measured.

Reversing the options changed the answer. Same essay, same six subjects in the other order: the subject flipped for between 7 and 16 of the 33 essays, a fifth to a half, depending on the checkpoint and how much of the essay it read. Hume found Jev order sensitive too: reversed options lifted a technical-support classification from the 0.84 band to the 0.93 band, which can carry an answer across a threshold without flipping the label.

Adding an option rewrote the answers to the others. Hume also found that Jev's options are not scored one at a time: append an irrelevant option and the odds between two options you never touched move. I repeated the test on two configurations, the fine-tuned checkpoint reading whole essays and the English checkpoint as shipped, sending the subject question four ways: the six subjects, the same six again as a control, the six plus a gardening option no essay here belongs to, and the six plus an exit option. The control came back identical on all 33 essays. Gardening was never once chosen, never rose above 0.06 of the probability, and still changed which subject won for three essays on the fine-tuned checkpoint and four on the English one. The exit option changed four and six. The confidence value moved on every essay, by up to a quarter of its range, and that is the number a threshold reads. So the exit option everyone recommends, me included, is a new question rather than a safety net, and a threshold set before it arrives was set on a different menu.

None of this says Jev would do the same. Laya's own card calls it "a fast base to specialise, not a zero-shot decision engine," and one team's writeup, on six hundred held-out items of its own, has Jev at 0.77, base Laya at 0.72 and a full fine-tune of Laya at 0.90. What the run does say is that the pilot works as a pilot. The reversal test alone would have stopped this branch from shipping, and the bar I set before the numbers was not met on any question. The go or no-go answer for this workload is no, and the filing stays by hand.

What it does not move

A repeated judgment with an answer space you can name is the cheap half of judgment, and Jev makes it close to free: ten thousand decisions over a thousand tokens each come to forty-two cents. Writing the question is the other half, and this blog has spent a year arguing that the question is where the work went. The criteria, the exit option and the threshold that turns a 0.52 into an action are somebody deciding what an error is worth, and Jev consumes those decisions rather than making them. The person whose ticket went to billing when it should have gone to technical is still a person, with an evening, who does not care that the mistake was schema-valid.

Jev answers in probabilities. The question is still yours. So is the threshold, and so is the answer when the threshold was wrong. A decision, it turns out, can be bought for four thousandths of a cent. Telling whether it was the right decision still cannot.