Decision Models (System One / Jev)

Overview

A decision model takes a piece of state (text, or JSON) plus a set of typed questions, and returns one typed answer per question with calibrated probabilities — in a single forward pass. It never generates text, never reasons in the open, and never returns a value outside the options you gave it. There is no completion to parse, no JSON to repair, and effectively no prompt-injection surface in the output: the shape is the API contract.

This is a distinct model category, not a provider. TypeSafe AI's Jev created it ("System One" models); Jev is the reference implementation, and the /v1/systemone wire format is now served by a whole ecosystem — OpenRouter, Ollama, llama.cpp, vLLM, SGLang, ONNX runtimes, and dozens of open checkpoints (see references/where-to-get.md).

This space is moving weekly — check what you already have first. New providers add decision-model endpoints as we speak, and inference platforms that never used the word a month ago are quietly exposing a decisions / systemone / classify route. Before shopping for a model, look at the AI provider you already pay for: it may already ship one, or a similar typed-decision concept, and you'd only need to change a base URL. If it doesn't have the word "decision" anywhere, ask whether it has a constrained/structured classification endpoint — that's usually the same idea wearing a different name. (Concretely: OpenRouter's default model list omits decision models; they only appear when you filter by output modality — ?output_modalities=decisions — which is how its site shows the Decisions tab.)

The mental model: you are not asking a model to write a decision, you are asking it to score pre-defined options. Milliseconds on a CPU, sub-cent on an API.

The three primitives (this is the whole API)