Many uses of a language model in business code are not conversations but decisions: is this invoice a duplicate, which category does this ticket belong to, how urgent is this message. Asking a chat model and parsing its prose is fragile. This release adds decisions as a model kind of their own.
The model kind
- Three question types Yes or no, a choice among named options, or a score on ordered levels. Each answer carries a probability.
- One input, many questions A call takes a text or any object, sent as JSON, and a list of questions, and returns the answers keyed by question name, in question order, from a single request.
- Validated before sending A malformed question is refused before any request, so it costs no quota and writes no audit row.
- Refusals typed A question the provider declines to answer comes back as a refusal, not as an empty or guessed value.
Providers and access
- Two providers OpenAI and TypeSafe, configured like every other model kind, through the user, default and global model resolution.
- From scripts Server scripts build questions from plain JavaScript objects and arrays and call a decision model from the same AI namespace as chat and embedding models, through the default, user, global or provider-named factory.
A decision returns a typed value a program can branch on, under the same pricing, quota and audit as any other model call.