A model that returns a decision, not text

If you haven't read about it yet, TypeSafe AI launched an interesting model on Tuesday. They emerged from stealth and announced early access.

Their model, Jev, doesn't generate text. You define the output type up front, hand it unstructured input, and it returns a typed decision with a calibrated probability.

What really interests me is RLCD, reinforcement learning from calibrated decisions, where the goal isn't just getting the answer right but knowing when it might be wrong.

As they put it: A model that is right 95% of the time but doesn't know when it's in the 5% can’t automate anything. A calibrated model turns human-in-the-loop into human-on-the-exception-path.

That feels especially useful as a calibrated judge in front of a tool call, deciding whether an agent's action goes through or not, and as a guardrail that returns a number I can set a threshold on, log, and defend in an audit. We don't always need faster agents. We need agents that know when to stop and hand off to a human.

There are still open questions. RLCD isn't published yet, and the benchmarks are in-house. Whether the probabilities hold on a real production distribution is something every team will have to prove with its own data.

But I'm excited to follow Jev's progress and explore how it could fit into our work.

Typesafe Jev: typesafe.ai/blog/introducing-system-one-models-and-jev