Designing a Robust LLM-Based Evaluation System for Agentic AI in Drug Discovery Through Human Alignment