Act Through the Gap: A Paired Execution Audit of Prepared Control Under Missing Feedback

A missed observation does not stop a physical system. Unless another command has already been prepared, the actuator continues to apply the last action and an uncertain feedback delay becomes an interval of stale open-loop control. We isolate the value of prepared execution with a paired experiment in which the policy checkpoint, physics, resets, and timing stream are held fixed. The only intervention occurs after the first action: an actuator-side buffer copies that action exactly and, if feedback is late, advances through a learned suffix that was computed before the gap duration was known. Across ten training seeds on Walker2d, Hopper, and HalfCheetah, seven timing processes, and hidden-physics shifts, the buffer improves validation robust mean by 7.1–13.9% and burst-and-drop return by 26.1–52.6%, while clean trajectories remain identical at every step. The gains persist on a disjoint held-out physics panel. A matched frozen-first-action ACT-style comparator is competitive but not uniformly better, and a three-seed Pusher diagnostic shows no aggregate gain. Finally, two disjoint five-seed hazard cohorts reveal a limit: at the highest tested gap rate, five of six task–cohort curves have fallen by at least 10% from an interior peak. The same turnover appears on fixed-horizon HalfCheetah, where early termination cannot explain it. Short prepared plans can bridge intermittent feedback gaps, but they do not replace sustained closed-loop control.