The controls tie the gains to the diagnosis

Every run scored as gain over its own base, in units of that source's spread (95% bootstrap intervals).

• Random SDF, same generator with the diagnosis withheld, is harmful more often than not.
• Adding more of a base's own real data never clears the noise band (max +0.83 SEM); full-size corpora 5–26× larger gain ≤ 1.46, and SPair loses 1.67 SEM / 4.6 GEO at scale.
• Routing the identical diagnosis to retrieval from 40k FlyingThings pairs never separates from a random draw: the needed cells barely exist in the corpus.
• Deformation buys semantics (+0.42) for a small geometric trade (−0.10).

Diagnosis is what makes motion designable.