How do you design deterministic evaluation rubrics to grade AI agent performance without relying on subjective or open-ended judgment calls?
