Decision Models

A decision model takes a piece of state and a set of typed questions, and returns a probability for each answer the caller supplied. It produces no text. Removing generation buys three properties an LLM cannot offer together: the answer is always one of the permitted options, every answer carries a calibrated confidence, and all questions about the same state are answered in a single forward pass, in tens to hundreds of milliseconds.

TypeSafe AI named the class "System One" models when it launched Jev on 15 September 2026, borrowing Kahneman's split between fast, intuitive judgement and slow deliberation. Within a month Kev, Strands Decider, Cloudflare, OpenAI, Microsoft and Liquid AI had all shipped models with the same interface. The framing is the one behind dual-process agent design, with the fast loop now served by a model built for it.

The class is defined by its interface, not its internals. Four constructions produce it.

Whichever route, the interface differs from an LLM asked to reply in one word on every row below.

AWS's open source entry, released under Strands Labs with weights on Hugging Face and the training data, scripts and every version's changes in the repository. It takes a Qwen3.5-2B-Base torso, removes the language-model head, and replaces it with a pointer head of about a million parameters. The head compares the hidden state at an <answer> position with the hidden state at each option's last token, and a masked softmax over those scores serves all three question types. The released checkpoint is v19; an earlier slot-head design performed significantly worse.

The project preregisters each training run: predictions and failure conditions are written before training and outcomes appended without editing them. Six retrains of one recipe varied by a standard deviation of 3.2 tasks, so the maintainers treat a difference under about 10 tasks between two single runs as unresolved.

The choice is mostly a deployment one. Jev for the strongest model behind an API, Strands Decider for a small machine or offline, Kev for the open recipe scaled up to a datacentre GPU.

JevBench, maintained independently by Benchmark Heaven, is the shared yardstick for the class. Each task gives a model state and a bounded rubric and scores the typed answer. Tasks are tiered from easy (explicit facts, clear intent) through standard and judge to hard, with the hard tier written by frontier LLMs.

The composite combines the four axes by harmonic mean, so a weak axis cannot be offset by a strong one. Cost and speed carry equal weight with accuracy, which is how Imajev-4B and Plumb-4B ranked above Jev 1.13 on the September board before the v1.6.1 re-measurement moved them down. Scores are not comparable across method versions. Two other indices exist, the Jev Decision Index on Hugging Face and the Artificial Analysis Decision Index, and the three disagree on a leader.

The Strands Decider repository ships a worked example that gates a tool call. An agent with a weather tool is prompted to be eager, so asked "What's the weather?" with no location it guesses a city. Before the call runs, an intervention handler asks the decision model two noul questions about the conversation and the proposed call, then maps the probabilities to an action.

1. 1. The agent proposes a tool call. The before_tool_call intervention fires before the tool executes.
2. 2. The handler sends the conversation and the proposed call as state, with two noul questions: are the argument values grounded in facts the user provided, and is it premature to call the tool before clarifying.
3. 3. The decision model returns a probability for each, locally, in one pass.
4. 4. A threshold turns the probabilities into an action. Ungrounded arguments return Guide rather than Deny.
5. 5. The model gets its turn back with the feedback and asks the user which city they meant.

Strands interventions return one of four typed actions. The handler is plain Python with no opinion about its contents, so a decision model, a Cedar policy or another agent can sit behind the same interface. Equivalent hooks exist around the model call and around the whole invocation.

The questions, threshold and policy in the example were chosen by hand. The point it makes is about cost: a check that completes locally in a couple of hundred milliseconds can sit on every tool call, where an LLM judge would add seconds to the step it guards.

A typed output guarantees the format of the answer, not its correctness. Vendor claims that the model cannot hallucinate refer to the first.

Related: Fast and Slow Brain for the System 1 and System 2 split these models serve, Judge and Escalation for the LLM judges a decision model can replace on simple verdicts, Guardrails and Stream Safeguards for the runtime checks they make affordable, and Steering for what a Guide action does to a run in flight.