Authored specification
Define NOUL questions with instructions and optional paired yes/no criteria, choice questions with declaration-ordered options, and score questions with ordered levels. The compiler derives result fields from the declaration; retain the specification revision and language with every evaluation. Display wording may be localized, but inference uses the frozen source-language specification.
A result contains specification_revision, actual model identity, input/output token accounting and each typed answer. NOUL yields a probability; choice yields the selected option, option distribution and confidence; score yields a value, ordered level distribution and confidence. These outputs are estimates, not authorization, verified facts or automatic state transitions. Apply an outcome only through a separately authorized operation with human review where the business policy requires it. See reviewed AI actions.
Source and derived interface
This Given fragment defines a static question set, requiring a containing source owner:
judgment Triage version=1
relevant noul "Does this concern the current request?" yes="Directly relevant." no="Unrelated."
route choice "Choose the destination." {support="Support",sales="Sales"}
urgency score "Assess urgency." [routine="Routine",soon="Soon",immediate="Immediate"]
NOUL yes and no are both present or both absent. A static judgment derives evaluate(state:text) -> Triage, the immutable Triage.specification:std.JudgmentSpec, choice enum Triage.route.choice and ordered level enum Triage.urgency.level. Import an exported judgment with the ordinary bound use form and invoke its derived evaluate operation through send ... as attempt; do not redeclare an ai.SystemOneV1 capability around it.
Runtime-choice questions use the accepted options=runtime form, optionally beside authored fixed choices. Such a declaration derives Name.options, evaluate(state:text,options:Name.options) and the pure Name.specification(options) function; request-local candidates use checked text IDs rather than fabricated dynamic enums. Static declarations retain their constant and original evaluate signature.
State is limited to 40,000 Unicode scalars. The portable profile allows 1–32 questions, 2–26 choice options, 2–10 score levels, at most 2,000 scalars per instruction/criterion and 24,000 combined specification scalars. Unsupported capacity fails before invocation, without truncation. Results must contain each question exactly once, exact option/level identities, bounded finite decimals and nonnegative integer usage. Distribution sums, selected maximum and score expectation use tolerance 0.000001; score is the weighted zero-based level index in [0,N-1], not a chosen level. Invalid or partial answers fail the delivery. Confidence has no guaranteed calibration.