A model or system that screens text before or after generation to block unsafe, harmful, or policy-violating content.