Left, top to bottom on one axis of the coverage table (magnitude by default, 12 log bands from 3 to 300 px in a 512-px frame; or the 16 direction sectors; or the 3×3 placement lattice): the training set's own annotated motion; the gap, the mass needed to lift every cell that sits below 5% of a uniform spread up to that floor (a small number of vectors, but its shape is the diagnosis); and where each arm's added pairs actually put their mass, from the pairs actually trained on at the prescribed dose, with the gap ghosted behind them. The prescription mirrors the gap; Random SDF piles mass at 100–300 px whatever was missing; own-source and FlyingThings additions repeat their source's habits. Right: sources whose missing motion is small-magnitude gain the most from filling it. Switch the addition: the diagnosed arms (prescription, and the real FlyingThings pairs the same diagnosis routed to retrieval) fall on the same falling line, so the fit is blind to whether the additions were real or rendered; Random SDF and Own Real, which carry no diagnosis, do not.