There is a benchmark-to-application gap between benchmark performance and real-world production performance, because standard evaluations are typically single-turn, short-context tasks with definitive answers.