A learned policy will eventually fail. The safety architecture around it — OOD detection, workspace limits, force monitoring, and emergency stops — is what keeps robots, objects, and humans safe when it does.