We advance a claim: learned (deep-learning and reinforcement-learning) control is replacingclassical robust control, H∞ synthesis and µ-synthesis, for the same reason deep convo-lutional networks displaced hand-engineered computer vision pipelines (SIFT, HOG, thedeformable parts model). Both are instances of one mechanism: replacing a fixed, pre-committedsignal/noisepartitionimposedbyahand-designedprocessinglayerwithalearnedpartition, thereby recovering predictive structure that the fixed partition discarded as noise.Theconservationlawsofcontrolarefundamentalandstayapplicable; thenewcontrollersarebound by them all the same. Our contribution is the cross-domain synthesis and preciselyseparating which limits are artifacts of an assumption and which are genuine information-theoretic converses. The argument is organized case by case: each claim is stated, givenits precise conditions and constants, attributed to its governing assumption, marked as stillconstraining or relaxed, and cited.