Jev is what picks the model.

Why calibration is the property that matters

Jev’s accuracy is roughly at parity with frontier models rather than ahead of them. That is not the interesting part. The interesting part is that its confidence scores track reality: independent testing after launch found that items it scored below 0.1 really did behave like the low-confidence cases, and items above 0.9 really did behave like the high-confidence ones.

That is precisely what a triage layer needs and precisely what an LLM judge asked to rate its own certainty does not give you. Our entire mechanic is gating expensive checks on a confidence number. Gate on a number that does not mean anything and you have built a random spend generator with good branding.

What this does not mean about hallucination

TypeSafe AI says Jev cannot hallucinate, and that claim is about Jev’s own typed decisions: a choice, a score, a calibrated yes or no, drawn from a fixed set of options it was given. It does not transfer to the free-text answer the model Jev picked goes on to write. That answer can still be wrong in the ordinary way any LLM answer can be wrong, which is the entire reason this product also runs a verification ladder underneath the routing. See what “no hallucination” can and cannot mean for a router for the longer version.

Why it is deliberately not the foundation

Jev is one implementation of a DecisionBackend interface, not a dependency this product is built around. Building a business on a single-vendor, recently-launched, waitlisted API is a risk we would rather not carry, and the vendor moving up-stack into routing and quality-gating themselves should be assumed rather than hoped against.

So the triage layer degrades instead of failing. If Jev is unavailable, slow or unaffordable, requests fall through to a conservative built-in heuristic that needs no network call and never throws. The heuristic is worse: it caps its own confidence at a level that reflects being a set of rules rather than a calibrated model, which makes it route upward more often and cost more. That is the right way for a fallback to be wrong.

What the receipt tells you

Every request names its own triage backend and confidence. You never have to guess which one decided a given call, and if a deployment loses its Jev key mid-month the receipts show exactly where the handover happened.