Jev: zero hallucinations means out of schema
Jev's zero hallucinations means zero out-of-schema outputs, not zero wrong answers. TypeSafe's own benchmark is ~68%, and the references are frontier models.
"Zero hallucinations" on Jev means zero out-of-schema outputs. Jev cannot return a value outside its declared schema. It can still be wrong. That is the single most misread claim about the model.
TypeSafe's introduction of System One models and Jev is the primary source for how the company states the claim. This page is the zero-hallucinations claim and its limits. It is not what a System One model is, not a typed platform-policy check, and not the cost and latency of routing every request.
Out of schema is the whole guarantee
Jev is not an LLM. It does not generate text. It takes unstructured program state and returns typed structured values plus calibrated probabilities and confidence scores, in a single parallel pass. TypeSafe calls it a System One Model: fast focused judgements, not deliberation, named after the fast/slow split in Thinking, Fast and Slow. DataCamp writes up that class. The guarantee applies to a fast judgement returned as data. String generation is given up entirely.
Out-of-schema means a returned value the declared schema does not allow. Jev cannot produce one. Zero hallucinations means zero out-of-schema outputs. A wrong value that the schema allows is still a legal return. The phrase means the value is in the schema. It does not mean the value is true.
The framing "mathematically cannot hallucinate" is a structural guarantee, not an empirical one. The structure is the declared schema. The return has to be one of the typed values the schema allows. The constraint forces that shape. It does not establish that the chosen value is true. Every value the schema allows remains available, including the wrong one.
Probabilities and confidence scores come back in that same pass. The pass is a fast focused judgement, not a deliberation, and the model does not generate text to explain the choice.
The accuracy number is ~68%
TypeSafe's own benchmark shows ~68% accuracy. The reference answers come from frontier models, not verified ground truth. Independent evaluation is still thin. Limited early access dates from 15 September 2026. "Zero hallucinations" with the accuracy line deleted is how the claim gets misread.
| What gets repeated | What the sources support |
|---|---|
| Zero hallucinations | Zero out-of-schema outputs. In-schema answers can still be wrong. |
| "Mathematically cannot hallucinate" | A structural guarantee, not an empirical one |
| Published accuracy | ~68% on TypeSafe's own benchmark |
| Reference answers | Frontier models, not verified ground truth |
| Outside checks | Independent evaluation is still thin |
~68% is TypeSafe's score on TypeSafe's benchmark. The reference answers were produced by frontier models, and they are not verified ground truth. Matching one is agreement with that model. It does not establish that the reference was checked against the world. A public entry for the model is the Wikipedia page.
A 50% probability means disregard it
Calibrated probabilities and confidence scores come back with the typed value. The schema constraint does not interpret them. A caller who stores the value and drops the probability and the confidence score has kept the field the structural guarantee covers and discarded the rest of the response.
Armin Ronacher, Earendil CTO, stated the rule he uses: "if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it."
A 50% probability obliges you to disregard the value. He calls that probability a possible coin toss, and the action he names is to disregard it. The typed value can be inside the schema and still be the result he will not use. Treating a legal value as a confirmed value is the misread, on a single call.
Jev is trained with RLCD (Reinforcement Learning for Calibrated Decisions), not RLHF or RLVR. The probability is part of the return, in the same pass as the value. At 50%, Ronacher's rule is to disregard it. The ~68% benchmark is a different fact: aggregate accuracy on TypeSafe's own benchmark, against frontier-model references, with independent evaluation still thin. One call's probability does not edit that benchmark, and the benchmark does not rescue a call that came back at a coin toss.
Diogo Almeida, the founder, is ex-OpenAI. He helped build ChatGPT and worked on RLHF. On the gap: "We have lightning in a bottle, and yet it is not useful." TechCrunch's 18 September 2026 report is the news account of the launch. A ban on out-of-schema values does not make the return useful. The return can still be wrong.
Classification is not a new task
KDnuggets states the corrective directly: "Classification is not new. Intent detection is not new. Zero-shot classification is not new." TypeSafe optimised an existing shape. It did not invent a new task.
A classifier aimed at a declared label set was already barred from answering outside that set. Zero out-of-schema outputs restates the constraint. It does not add a new kind of correctness. The remaining job is the one classification already had. Pick the right label. Drop it at 50% probability, the case Ronacher treats as a possible coin toss and disregards. TypeSafe's ~68% is the company's measurement of the pick, scored against frontier-model references. Ronacher's line is the drop.
The claim stops at the jobs Jev does not do. It cannot process images yet, so a picture is an input it cannot take. It gives up string generation entirely, so it will not produce a paragraph. It is not for reasoning, coding, or open-ended generation, so a derivation is outside the job. The structural guarantee says nothing about those tasks.
What the claim says
Hold the launch line to one parsing. Zero hallucinations means zero out-of-schema outputs, not zero wrong answers. The ban is structural, not empirical. TypeSafe's own benchmark is ~68% against frontier-model references, independent evaluation is still thin, and a 50% probability is a possible coin toss to disregard.