A specific safety boundary or checkpoint in medical AI that prevents harmful outputs, such as escalation protocols or dosing validation.