I argued for years in favor of keeping the minimum runtime of a number of samples; never found much support.