Last updated: September 2026
Can Jev hallucinate?
Jev cannot return a value outside the schema you declare. It can return the wrong value inside it. Everything written about Jev and hallucination collapses into that one distinction, and most of the coverage repeats only the first half.
"Zero hallucinations" is a claim about schema conformance, not about factual accuracy. TypeSafe's own launch post says of the 0% figure: "Our number is not empirical. Schema matching is guaranteed." Below: what the guarantee covers, a worked example of a confidently wrong answer, and how to catch one.
Two things get called hallucination. Jev eliminates one of them
A language model asked for structured output can fail twice: it can break the shape, and it can get the content wrong. Jev removes the first failure by construction. The second one is untouched.
The practical version: your parser will never throw, and your error handling still has to assume the answer can be wrong.
A worked example of a confidently wrong answer
Sarcasm is the cleanest case, because the literal reading and the intended reading point in opposite directions. Here is a Noul question asked about a sarcastic post. The request and response shapes are the real ones from TypeSafe's API reference; the values are a constructed illustration.
{
"state": "Absolutely love that the API went down during our launch. Perfect timing. Ten out of ten.",
"model": "jev-1.13.0",
"questions": {
"positive_about_product": {
"type": "noul",
"instructions": "The author is expressing a positive opinion about the product."
}
}
}{
"model": "jev-1.13.0",
"answers": {
"positive_about_product": {
"type": "noul",
"noul": 0.91
}
},
"usage": { "input_tokens": 41, "output_tokens": 2 }
}That answer is schema-valid. Every field is the one you declared, the float is inside 0 to 1, and no parser anywhere will complain. It is also wrong, and it is wrong at 0.91, which is above the threshold most people would set for acting automatically.
Two things make this specific failure worse than it looks. TypeSafe documents that Jev reads instructions and state literally, and says that negation and indirection degrade accuracy. And a Noul returns only a noul float. There is no confidence field on a Noul answer, so the number that is wrong is also the only number you get.
The same shape shows up in content scoring
The free post scorer is built on Jev, and one of its questions is a Noul for text that reads as AI written. The criteria name em dashes, triadic lists and uniformly long sentences. A person who happens to write that way gets a high value on that Noul. The answer is valid, the writer is human, and the score is wrong. That is why the tool shows the probability rather than a verdict.What TypeSafe and its critics actually said
This is not an outside correction that TypeSafe disputes. The company made the narrower claim itself, and then said so again when pushed.
- TypeSafe, in the launch post. On the 0% hallucination figure: "Our number is not empirical. Schema matching is guaranteed."
- TypeSafe's CEO, on Hacker News. In the launch thread, Diogo Almeida acknowledged that Jev can return a schema-valid answer that is factually wrong. His defence is the confidence score: low-confidence results are the ones you discard.
- The top Hacker News objection. "Sure, it can't emit an invalid type, but it can still emit a completely wrong valid value. You can enforce structured output from an LLM too."
- The Register. The hallucination-free claim "really isn't a fair comparison as its output is not natural language," since a structured response with probabilities does not preclude being incorrect.
- Sean Goedecke calls the immunity claim "a semantic dodge", on the grounds that the model still makes mistakes by selecting the wrong choice.
The only independent test so far is Mike Taylor's at Every, who ran 777 judgments in under 0.7 seconds for an estimated quarter of a cent. Jev caught 6 of 7 planted defects. Claude Fable 5.1 caught 7 of 7. Jev missed the same defect, an unexplained action, in all three runs. Taylor called it useful as an early warning system and said he would want a more thorough accuracy check before putting it into production.
Where Jev is documented to go wrong
TypeSafe publishes a model jaggedness page for jev-1.13 that lists the failure modes. It is unusually honest for a vendor doc and it is the best predictor of where your own integration will break.
The primitives themselves are covered in Choice vs Score vs Noul, including which answer types carry a confidence field and which do not.
How to catch a wrong answer
You cannot make a Jev answer correct. You can decide which answers get to act on their own.
- Gate on confidence, per action. Choice and Score return a
confidencefrom 0 to 1, computed from the probability distribution. TypeSafe's guidance is to act automatically above roughly 0.9 for high-stakes actions and route below roughly 0.5 to a human, and it refuses to give universal numbers. Gate a delete harder than a label. - Remember that a Noul has no confidence. Threshold the
noulfloat yourself, and keep a dead band in the middle where nothing happens automatically. - Read the full distribution, not the top answer. A Choice returns
probabilitiesover every option. Two options at 0.44 and 0.42 is a coin flip wearing a label. - Keep arithmetic out of it. Count, compare dates and sum in code. The model does the judgment, your code does the maths.
- Pin the version. Log
jev-1.13.0rather thanjev-latest, or the thresholds you tuned will drift the day the alias moves. See the model strings and limits. - Run it in shadow first. Log answers for a week without changing behaviour, then plot confidence against accuracy on your own labels and set thresholds from that.
None of this is unique to Jev. What is different is that the failure is always a plausible value rather than a crash, so nothing in your stack notices it for you.
Frequently asked questions
Can Jev hallucinate?
It depends which kind of hallucination you mean. Jev cannot return a value outside the answer space you declared, so it never invents a field, a type, or an option that does not exist. It can still return a schema-valid answer that is factually wrong. "Zero hallucinations" is a claim about schema conformance, not about accuracy.
Does TypeSafe admit that Jev can be wrong?
Yes. TypeSafe's launch post says of the 0% figure: "Our number is not empirical. Schema matching is guaranteed." On Hacker News, TypeSafe's CEO acknowledged that the model can return a schema-valid answer that is factually wrong, and argued that confidence scores let you discard the low-confidence ones.
What does "zero hallucinations" actually guarantee?
That every answer is inside the answer space you declared. A Choice returns one of your option keys and nothing else. A Score returns a float inside your level range. A Noul returns a float from 0 to 1. Your parser will never throw. That is the whole guarantee.
How do I catch a wrong Jev answer?
Read the confidence, not just the answer. Choice and Score return a confidence from 0 to 1 derived from the probability distribution. TypeSafe's docs suggest acting automatically above roughly 0.9 for high-stakes actions and routing anything below roughly 0.5 to a human, while refusing to give universal numbers. A Noul has no confidence field, so on a Noul the probability itself is the only signal.
Is Jev more accurate than a large language model?
No. On TypeSafe's own four-workflow eval, Jev scores 67.8% against GPT-5.6 Terra at 67.9%, with Sol at 74.1% and Opus 5 at 73.1% both ahead of it. The pitch is comparable accuracy at 0.4 seconds and $0.0004 per case, not higher accuracy. The eval also has no human ground truth: the reference answers are the average of two other models.
Keep exploring
The rest of the Jev reference, and the tool it runs behind.
Is Jev an LLM?
No. It returns typed decisions with probabilities and generates no text at all.
Choice vs Score vs Noul
The three primitives, with the exact request and response JSON for each and the gotchas.
Jev rate limits and context window
64k tokens per request, 1,200 requests per minute, and what that means at social-feed volume.
How to get Jev API access
The waitlist, plus the three gateways that do not need one.
Jev vs GPT
The four-workflow benchmark as a table, and what each model cannot do.
Will it go viral?
A free post scorer that runs on Jev in production. No signup.
See a Jev answer with its distribution attached
Will it go viral? scores an X draft with seven Jev questions and shows the confidence band, because the distribution is what the model actually returns. It is free and needs no signup. If you want to build something similar on top of publishing, the OpenTweet API is the part that ships the post.