What "cannot hallucinate" actually guarantees

The guarantee#

Jev's output space is the type you declared. A Choice over three options cannot return a fourth; a Score over five levels cannot return level six; a Noul cannot return a string. There is no decoding step where a wrong token could slip through, because there is no decoding step. TypeSafe describes violating it as "mathematically impossible", and that is not hyperbole — it is a property of how the output is constructed.

This is worth something real. It removes an entire class of production failure: the parse error, the retry loop around json.loads, the enum value that was almost right, the tool call naming a tool that does not exist. TypeSafe's framing of why that matters is fair — a hallucinated tool call is an inconvenience in a chat agent and a deal-breaker inside a system with latency guarantees or several layers of dependency.

What it does not guarantee#

It says nothing at all about whether the returned value is correct. The model can pick the wrong option, misread the evidence, or assign a high probability to a bad answer — and every one of those is type-safe.

A Hacker News commenter put the same point as a question nobody answered well: "if it puts a high confidence value on a wrong answer, that's still hallucinating, no?" And The Register's assessment was blunter — calling the model hallucination-free "really isn't a fair comparison as its output is not natural language... that does not preclude the possibility of being incorrect."

The 0% on the chart is an axiom#

TypeSafe's launch post includes a hallucination-rate chart with Jev at 0%. Their own footnote explains where that number came from:

So the bar is not a measurement. It is a definition, plotted next to measurements of a different thing — the LLM bars are empirical schema-failure rates from OpenRouter traffic, which TypeSafe also notes carries routing bias. Comparing a theorem to a sample on the same axis is the part that does not hold up.

Why the distinction has teeth#

Because of where this model is meant to sit. Every deployment pattern in these tutorials puts Jev in a control path: gating a tool call, approving a transaction, screening content, routing a ticket. In a chat interface a wrong answer is visible to a human who can push back. In a control path it is a branch nobody watched.

Combine that with the documented fact that state is not treated as hostile, and the risk profile is clear enough: text an attacker controls, feeding a decision nobody reviews, made by a model that cannot return an invalid answer but can certainly return the attacker's preferred one.

• Never let a single Jev answer be the last thing between untrusted input and an irreversible action.
• Give every question an honest exit — other, none of the above, insufficient evidence — because the probability mass has to land somewhere, and without an escape route it lands on the nearest wrong option with real confidence.
• Keep consequences under code control: permissions, limits, idempotency, rollback.
• Treat a type guarantee as removing a class of crashes, not a class of mistakes.

In short#