As large language models become more capable, users are tempted to delegate knowledge tasks where models process documents on their behalf and provide the finished results. But how far can you trust ...