What a decision model is

A language model answers by generating text, one token at a time. To get a decision out of it you parse that text and hope it matches one of your options. It might not, and any probability it quotes is itself generated text.

A decision model takes a state (the situation, as text or JSON) and typed questions, and returns an answer per question in a single pass. TypeSafe calls this class a System One model, after the fast, intuitive thinking in Daniel Kahneman's Thinking, Fast and Slow. The answer can never fall outside the options you listed, so there is nothing to parse. It can still be the wrong option.

choice

Pick one of named options; returns a probability for each. Every game here uses one: a move, a lane, hit or stick.

score

A position on a scale you describe, such as severity 0 to 3. Returns the expected level and the spread.

noul

Is this true? Returns a single probability. Used for yes-or-no checks: is this spam, is this urgent, is this a jailbreak.

Because every answer carries probabilities, code can act on the confident ones and send uncertain ones to a person or a larger model. That is the pitch: fast, cheap, typed decisions inside ordinary software.

One decision from a game on this site: what the model was sent, and exactly what it answered. Nothing is edited, and the whole game is replayable in the Arena.

Neither model sees pixels or coordinates. Each game turns its state into short sentences, because decision models read meaning and are poor at arithmetic. Both models get exactly the same text and the same options.

Jev, the closed model that started this. Laya, the open-source answer you can run yourself. Side by side, with what we measured.

Sources: TypeSafe's launch post, Laya on GitHub, a deep dive into Jev.