Actually, my takeaway from this paper was exactly the opposite: if you understand how floats work, you can trust the results of floating-point calculation.