The growing gap between benchmark scores and actual model capabilities due to stale verifiers, infrastructure changes, and accumulated defects in benchmark tasks.