SayWhen: Set-Valued Semantic Commitment for Low-Regret Early Speech

Speaking early turns a prediction into an irreversible public commitment. Committing to one predicted ending may require repair, whereas waiting forfeits conversational overlap. We introduce SayWhen, a test-time controller for typed, set-valued pre-endpoint commitments. Its model path derives task-directed acts shared across typed continuation plans from a pretrained language model; a deterministic retail schema covers predefined order workflows; and a persistent release gate withholds a candidate when its detector flags a loss of support before playback. We formalize regret-free early speech (RFES), which requires a supported task-directed release before the endpoint and no scored release incompatible with the realized interaction. In a controlled branching diagnostic with generated plans and timing fixed, the typed frontier reduces observed mean public-commitment regret from 0.068 to zero, with RFES 2.2 percentage points below the Matched Top-One baseline. In a fixed 30-task $\tau$-Voice full-duplex simulation, semantic veto begins eligible task-directed audio an average of 1.51 seconds before the user-audio endpoint among tasks with early speech, while both semantic veto and endpoint waiting pass 16 of 30 tasks. These results suggest a broader view of real-time responsiveness that considers both when an agent speaks and which public commitments are supported by the available evidence.