Explore how statistics can make AI evaluation more rigorous. This dissertation defense presents three novel frameworks for testing large language models with scientific precision, from prompt sensitivity analysis to bridging human and automated judgments.