The Model Loop: Recommend, Act, Evaluate (and Show Your Work)

Walk through inference as a loop: generate recommended primitives plus grounded reasoning, then evaluate the utterance with scoring heads conditioned on the prior advice. Emphasize transparency: the system cites evidence and explains why.

At inference time, PRISM operates as a closed loop. Given the current dialogue state and speaker–audience profiles, the recommendation module generates a set of primitive suggestions plus grounded reasoning that cites specific evidence spans. The speaker then produces an utterance. The evaluation module scores that utterance against the recommendation, using scoring heads that are conditioned on both the recommendation and the context.

The transparency requirement is functional, not decorative. A recommendation system that cannot explain why it chose a particular strategy for a particular audience in a particular moment cannot be validated, debugged, or trusted in high-stakes communication contexts. Grounded explanations make the system auditable.