I like Mark's phrasing as well. For precision, I'd still like to give an exact algorithmic formulation of what "large enough" means in this context, though.

Something like:

    Each bitwise operation has the same result as though carried out in two's complement using a bit-width that's large enough to represent the inputs. ("Large enough" for this purpose is ``1 + max(x.bit_length(), y
.bit_length()``, with the extra bit being needed to handle sign extension appropriately)