Drug discovery models fail on new chemical scaffolds because they learn patterns in training data, not chemistry. Causal-Phys is a graph neural network architecture built around a single constraint: predictions must be grounded in atoms that causally drive biological activity, verified against both molecular topology and three-dimensional geometry. The target is a model that generalizes to unseen scaffolds not by having seen enough examples, but by having learned something real. If it works, the implications for early-stage drug discovery extend well past any single benchmark.