> it must return the same results everywhere.

We solved that for floats like 10 years ago. Let alone the fact there are better formats, like posits, or fixed point numbers, that also solve this problem very easily.