Glossary

Shared vocabulary

Terms used across this guide. Prefer explicit values such as “not yet tested” over ambiguous blanks in evidence fields.

State

The structured evidence the orchestrator assembles for a decision call—only the fields each question needs, retrieved under the caller’s permissions, with freshness rules.

Typed output

Constrained answer shapes defined in advance (Choice, Score, Noul) with probability information. Schema guarantees prevent malformed shapes; they do not make the judgment factually correct.

Choice

Selects one option from a set you define and returns a distribution across options. Useful for routing and classification when categories are agreed.

Score

Places the state on an ordered rubric you describe. Treat as a signal for queue ordering; keep contractual obligations in ordinary rules.

Noul

Returns the probability that a clearly defined yes/no condition holds. Name a condition a reviewer could verify.

Calibration

Asks whether groups of predictions behave as their probabilities suggest. It does not guarantee that an individual answer is correct.

Abstention

A designed “ask a person” or review outcome when evidence is insufficient, confidence fails the gate, or the action is irreversible.

Idempotency

Ensuring that duplicate delivery or retries produce one business outcome. Use a durable unique key; separate prediction from consequential actions.

Policy

Deterministic code that turns validated model answers into authorised outcomes—thresholds, allowlists, reversibility checks and reason codes.

Confidence

A summary of how probability mass is spread. A clue for the policy gate—not permission to act.

System One

TypeSafe’s category of models for short, bounded, software-facing decisions. Jev is the first public model in that category.