Lesson two of a course on holding the model while an AI does the work. The delay is Donella Meadows’s (Thinking in Systems, 2008): a balancing loop whose evidence arrives late overshoots, because the hand keeps acting on a level that has not yet felt the last action. The correction is to act once, then wait. In practice the lag sits in the evidence (a test run, a review, a deploy), not in the plumbing.