Overview
A cyber-physical security study on the ROAD dataset (Oak Ridge National Laboratory): 33 verified attack captures including 13 masquerade and 12 ambient drives. The work recovered the signal map from raw DBC mismatches, built a physics taxonomy of all 26 injection captures, where every class except reverse-light violates at least one kinematic consistency law with ratios of 178-1273×, and showed physics plus range statistics catch 3 of 4 classes with recall 1.0 at a 0.1% ambient false-positive rate, detecting within a single CAN frame. Milestone four reports an honest negative: the PINN twin shows parity with the physics-free model across four data regimes and three loss weights, so attribution credit goes to the physics, not the neural network. All five pre-registered milestones are complete and a five-page paper draft is compiled.