Rigorous controlled evaluations typically require methodological safeguards such as fixed prompts, multiple independent runs averaged together, and blind evaluation.