Tag
This paper derives sharp theoretical limits for honest uncertainty in repeated evaluation under a hard budget, with applications to language model and agent benchmarking, demonstrating practical improvements in interval width and MSE.