Definition
In current AI usage, a decision model evaluates supplied context and returns bounded judgments that application code can consume directly. A judgment may select from supplied options, assign a score, or estimate the probability that a condition holds. The application defines the question and answer space, then decides what action the result permits.
A support system might ask whether a message needs escalation, choose its destination queue, and score its urgency. Those are separate judgments. A high escalation probability is an input to policy, not authorization to send a message or change an account.
Origin and current usage
TypeSafe AI introduced Jev on September 15, 2026 as its first System One model. Founder Diogo Almeida's announcement described typed probabilistic decisions for software. Credit for that product belongs to TypeSafe; the announcement does not establish the first coinage of decision model.
In his September 21 interview, Almeida explained that TypeSafe prefers System One for a broader machine-native model class. Its current primitives are decisions, but he envisages other output types. System One and decision model therefore should not be treated as universal synonyms.
OpenAI released its Decisions API beta on October 6, 2026, powered by GPT-6 Luna. Its predicate, choice, and score interface is another example of bounded decision output. An API release alone does not establish a new model architecture.
Scope and terminology
TypeSafe's broader System One proposal and OpenAI's Decisions interface describe different scopes. This working definition covers bounded judgments consumed by software. The output contract does not require a particular neural architecture or settle whether a model reasons internally.
Operational significance
Evaluate accuracy and calibration on the actual question, answer choices, and operating distribution. Set action and escalation thresholds using the consequences of errors. A well-calibrated group of predictions can contain wrong individual answers.
Typed output constrains representation; it does not make a semantic judgment infallible. TypeSafe's Jev 1.13 documentation records difficulties with arithmetic, date comparisons, indirection, adversarial input, and option ordering. Capabilities and limitations need version-specific evaluation.
Distinguish it from nearby terms
- Classification assigns labels. A decision interface may additionally expose scores or probabilities, while classification remains one of its underlying tasks.
- A reasoning model allocates inference effort to multi-step problems. Decision model describes a use and output contract without prescribing one universal architecture.
- Structured output guarantees a supported output shape, not that a chosen option is correct.
- Function calling proposes a tool invocation. A decision can select a tool, but execution and permission checks belong to the application.
Check your understanding
A model assigns a refund request a 0.94 probability of eligibility. Which additional evidence and authorization must the application check before issuing the refund, and how would you test whether that probability is useful?