But the errors aren't random; they are (experiment shows) overwhelmingly in the direction of optimism. How can it take lots and lots of expensive little detail

But the errors aren't random; they are (experiment shows) overwhelmingly in the direction of optimism. How can it take lots and lots of expensive little detail work to correct such a gross statistical bias with a known direction?