AI researchers and labs have advanced by leaps and bounds in evaluating AI models for everything from safety and compliance to sycophancy and alignment. But it appears companies and developers are faced with a new, specific need: making sure that their AI system behaves as intended for their specific product or service.