Prompt-only meeting delegates stay silent through 51.4% of their turns.

CAPA measures LLM agents delegated to attend a meeting on someone's behalf, and prompt-only versions miss half the absent participant's talking opportunities on the AMI corpus. The fix is an architecture, not a better prompt: separate Perceiver, Predictor, Controller, Generator, plus two judges and a recalibrator scoring the forecast against the next observed turn. Silence drops to 2.5%, credited recovery goes 26.1 to 52.2, hallucination holds at 0.6%, judges align with human annotation at kappa 0.71. The general shape, split "should I act" from "what should I say" into separate components, transfers to any agent that has to decide when to interject.