Read this before you trust the phrase
What a “no hallucination” router can honestly claim.
People searching this phrase want to know whether a product can promise them a model will never make something up. No credible product can promise that about a model’s free-text output, and we are not going to be the one that pretends otherwise. What follows is the precise, narrower claim we can actually back, and the specific checks that back it.
Live on this deployment
Jev is live here as the routing decision layer, and the verification ladder described below runs on every request regardless of which backend triaged it.
The claim, stated precisely
TypeSafe AI says Jev cannot hallucinate. That claim is about Jev’s own typed output only: a choice from a fixed set of options, a score, a calibrated yes or no. There is no free-text generation step in a Jev call for a hallucination to occur in. Judged against that narrow, well-defined task, the claim is reasonable, and it is not the claim most people searching for a “no hallucination router” are actually asking about.
What people actually want to know is whether the sentence a model writes back to them, the free-text answer to their question, can be trusted not to contain something invented. Jev’s guarantee does not reach that far, and no one else’s does either, for a structural reason: once a model is generating open-ended text rather than picking from a fixed set of options, there is no closed set of possible outputs left to be calibrated over. The same model that Jev correctly and confidently routed a request to can still write a wrong sentence. This does not eliminate hallucination risk in that sentence, and nothing about the routing decision that preceded it changes that.
Why this is not a loophole, it is the reason we built a second system
If a calibrated router were enough on its own, there would be no reason for llm11 to run anything after generation. We run a verification ladder on every answer specifically because routing correctly and answering correctly are different problems, and solving the first does not solve the second. The ladder does not know or care which backend triaged the request; it checks the answer that came back, the same way, every time.
What the ladder actually checks
| Rung | When it runs | What it catches |
|---|---|---|
| schema | Always, on every request. | Malformed output: JSON that does not parse when JSON was asked for, a response that ignores a length or format constraint, an empty answer. |
| groundedness | When the caller supplied context. | A factual claim in the answer that the supplied context does not support. Every claim gets held against the source you gave us, not against the model’s general knowledge. |
| cross_model | When triage flags the request as high stakes or worth a second opinion. | Confident fabrication a single model won’t catch in itself. A different model, from a different provider, answers the same question independently, and a material disagreement in fact, number, or conclusion fails the check. |
| self_consistency | Reserved for high criticality combined with genuine triage uncertainty. | An answer that will not hold still. The same model is resampled and the spread across runs is scored; instability correlates with the model being unsure even where the answer sounds certain. |
Fail any rung and the request escalates automatically, once, to the strongest remaining model in the candidate pool, which then gets re-checked before it reaches you. See how the routing pipeline and the ladder fit together for the full sequence.
What none of this catches
Be clear about the limits, because a vague promise here is worse than a specific one. A free-floating factual error on a question with no supplied context has nothing for groundedness to check against, and cross-model only helps if the second model happens to know better rather than sharing the same wrong belief. An answer that is faithful to context you supplied but where that context was itself wrong will pass groundedness, because groundedness checks faithfulness to your source, not the truth of your source. None of this promises that a model’s free-text answer will come back free of hallucination, and we do not claim otherwise anywhere on this site.
What we do claim is narrower and checkable: a defined, describable set of failure modes gets caught before an answer reaches you, each one by a named check whose logic is public, at a cost printed on the receipt. See the full account of what we catch and what gets past us for the complete, unhedged list.
The receipt is the proof, not the marketing copy
Every request comes back with a record of exactly what ran: which rungs fired, whether each passed, what triggered an escalation if one happened, and which backend made the routing decision in the first place. You do not have to take a claim about hallucination on trust, because the receipt names the specific check that either found a problem or did not, on that specific request. A claim you cannot audit is not a claim worth much, so we built the audit trail as the actual product rather than as a footnote to it.