<p>Artificial intelligence is entering materials science and manufacturing at a moment when retrieval is often mistaken for discovery. We argue that models trained to optimize plausibility within a closed theoretical system are structurally biased toward interpolation and away from the anomalies from which scientific revolutions arise. In materials and manufacturing, this limitation becomes acute because the design space is combinatorial, multiscale, and physically constrained. What is needed is not another assistant for summarizing what is already known, but systems that can assemble compositional world models, invert the forward problem, falsify their own principles, and test those principles against simulation and fabrication. Drawing on recent work in graph-native reasoning, inverse protein design, multi-agent scientific discovery, and AI-integrated manufacturing, we outline an architecture for discovery built on three coupled capabilities: world-model construction, adversarial falsification, and physical grounding. We make this architecture operational through a minimum description length gate: A proposed world-model break is accepted only when the revised model encodes the accumulated evidence in fewer bits than the model it replaces. We illustrate the loop with a residue-level protein mechanics case study in which a Breaker–Builder agent system revises a symbolic graph model of crystallographic B-factor data through accepted, rejected, and retracted hypotheses. The central claim is that materials innovation will accelerate when AI moves beyond retrieval and surrogate prediction toward systems that can generate, challenge, and physically realize new hypotheses. In that sense, discovery requires breaking the current world model in order to build the next one.</p>