Jev and the Case for Structured, Confident Outputs
SEPTEMBER 22, 2026 · LLM
I ran into TypeSafe AI's post introducing System One models and Jev this week and I'm still turning it over. The pitch: skip free-text generation, which can hallucinate, and have the model output a structured value with a calibrated confidence score attached. Their founder put the gap in plain terms: models have been "superhuman at chat for years," so where's all the automation? I think that's a fair question. A model that's great in a chat window isn't automatically useful glued into a pipeline.
Structured output beats a generated string
A normal LLM call in an agentic pipeline still hands you back a string, and you're trusting the model got the format, the types, and the facts right. Jev instead returns a structured value that can't have a type error, because of how it's built, plus a confidence score that's actually calibrated: higher confidence means higher accuracy, not the overconfident guessing most LLMs do. For anything gating a decision in a pipeline, that beats parsing JSON and hoping for the best.
RLCD instead of RLHF or RLVR
What makes this work is the training. Most LLMs today are tuned with RLHF, human preference, or RLVR, verified rewards on tasks with a checkable answer. TypeSafe is using something they call Reinforcement Learning for Calibrated Decisions, optimizing directly for honest probabilities instead of what a rater or reward model happens to like. You're not training the model to give answers people approve of. You're training it to know how sure it should be.
Why this matters outside of chat
The numbers are what make this feel practical instead of academic: 70-500ms end to end versus 3 to 329 seconds for frontier models on comparable tasks, and output tokens cheap enough to be free. If that holds up under real use, a calibrated structured decision at that speed and cost is a genuinely different tool than a chat completion. Useful anywhere you need a fast, trustworthy decision instead of a paragraph.
What's next
I've been watching people try Jev out this week, and I want to test it on something with a real right and wrong answer instead of something subjective. Chess seems like the move: legal moves, a clear evaluation, no arguing about whether the model was right. I want to see what a structured, confidence-scored model actually does when a move is either legal or it isn't.