generative-ai ai-agent validation system-design reliability closed-loop llm-as-judge first-principles eventual-consistency