Hugging Face has outlined a “decision model” interface for tasks where software needs a bounded choice or probability rather than generated prose.
A request supplies a state, such as a customer message, and named questions with predefined possible answers. The model returns scores or probability distributions for primitives including choosing among options, rating an ordered scale and estimating whether a statement is true. A support system could therefore route a ticket, rate urgency and detect a refund request without generating a free-form JSON answer token by token.
The approach overlaps with conventional and zero-shot classification. Unlike a fixed classifier, its candidate choices can change at inference time. The distinction is a consistent interface designed around several decision types and explicit probability distributions, potentially backed by models trained specifically for that role.
Probabilities are not guarantees of correctness, and the proposed interface does not eliminate the need for calibration, thresholds or human escalation. It also works only when developers can define the relevant choices in advance. For routing and other constrained workflows, however, removing free-form generation can simplify validation and make uncertainty easier for downstream software to handle.