I would suppose that the standard should simply abstain from standardizing evaluation involving expressions of mixed radices by requiring the programmer to explictly convert operands the way he wants. Numerical conversions between radices could be covered by the base conversion rules for input/output. NaNs are more interesting.