A real-world benchmark tests whether powerful AI models can keep an investigation trustworthy when new evidence invalidates their conclusions.