The idea is to make errors obvious rather than subtle. A NaN output is far more obvious than a number that is 2% off.