Has anyone put Jev in a real product yet, or is this still a waitlist toy?

TypeSafe launched Jev last week. It is a System One model meant for typed decisions inside software (classify, route, score, extract, branch), not chat.

I am not on the team. I am trying to figure out if this actually changes how products get built, or if it is just another structured output wrapper with better marketing.

The pitch in plain terms: you give it unstructured state, you define the output schema up front, you get calibrated probabilities back in tens or hundreds of milliseconds, and it is not supposed to spit out invalid types. Their argument is that chat models optimized for writeups are the wrong interface for automation.

That could matter if you already have LLM calls acting like fuzzy if statements. It is also the kind of claim that looks great in a demo and then gets weird in production.

What I want to hear from people here:

• Have you used early access, or are you planning to try it on a real workload?

• If you have used it, what did you point it at?

• Did it replace an existing model call or sit next to one?

• If you are not trying it, what is the blocker? Schema defined up front, no generation, trust, the waitlist, or JSON mode already being good enough?

The product question I keep coming back to: if decisions get this cheap and typed, do you strip intelligence out of a feature, or do you add more decision points you previously could not afford?

I care more about boring production paths (routing, moderation, extraction, guardrails on another model’s output) than the Doom and Wikiracing demos.

If you already swapped something in a live product, what broke first?