What Jev is, and how it might score your interview
On 15 September 2026 a company called TypeSafe AI released a model called Jev that cannot write a sentence. You give it a state and a question with a fixed set of possible answers, and it returns one of those answers with a confidence number, in under half a second.
That sounds unrelated to an interview until you notice what scoring actually is. Deciding which bucket an answer falls into is exactly the kind of question a model like this is built for.
What is Jev?
TypeSafe calls it a System One model, after the fast half of how people are said to think. It comes from Diogo Almeida, a former OpenAI researcher, and it works nothing like a chat model. You send it the facts of a situation plus a small set of questions with known possible answers, and it answers every one of them at once rather than writing out a reply word by word.
Each answer comes back typed, matched to a fixed list of options, with a probability attached. TypeSafe prices it at a fraction of a cent per case, and it responds in 70 to 500 milliseconds, against several seconds for a model that writes text.
How is that different from a chatbot?
| A chat model | A decision model like Jev | |
|---|---|---|
| What it gives back | A written answer | One option from a fixed list |
| How it answers | Word by word | All at once |
| Speed | A few seconds or more | Under half a second |
| What comes with the answer | A confidence you have to ask it to state | A number it was trained to calibrate |
| Good at | Explaining, writing, holding a conversation | Sorting something into a known category |
Both are still guesses. The difference is what shape the guess takes, and how fast and cheaply you can ask for one.
Why would an interview platform want something like this?
Scoring an answer is closer to sorting than to writing. Did the candidate address the question, give a real example, and reason clearly, or not. Those are checklist questions with a limited set of outcomes, which is precisely the shape a fast decision model handles well and a text model handles at far higher cost and latency.
None of this means a platform only needs one kind of model. Writing the actual feedback you read afterward, in full sentences that explain a judgement, is a different job, and that still calls for a model built to write. A serious platform likely uses more than one tool, not a single one for everything.
Does not being able to write mean it cannot get it wrong?
A model like this cannot hand back a malformed answer, because every answer has to come from the list it was given. It can still pick the wrong item on that list. The guarantee is about shape, not about whether the verdict was right.
TypeSafe's own published comparison measured this kind of model agreeing with other AI models roughly as often as those models agree with each other, not against a human answer key. Treat any claim about a model's judgement, including this one, as something to check rather than take on trust.
What should make you trust a score more, or less?
- A report that quotes your own words back to you as the reason for a mark, rather than a bare number.
- A score that comes with a plain explanation, in sentences, of what it was checking for.
- A way to see your own transcript afterward and compare it with the feedback.
- If a platform says nothing about how a mark was reached, that is worth asking about directly.
What this actually changes for you
Almost nothing about how you should answer. You still do best with a clear situation, what you did, and what happened, said in your own words. What changes is the honest picture of what is running behind the interview: increasingly, a mix of models built for different jobs rather than one model doing everything, and that is a reasonable way to build this kind of system, not a reason for suspicion on its own.