Safety enforcement applied at the moment an AI system takes action, separate from training or inference-time controls.