I'd assume that there's a difference between picking the best _token_ across an assortment of randomly selected tokens, versus picking the best _string_ of randomly-selected tokens.