It looks to me that following the ?: specification would actually be easier to follow and more correct.
I sketched an implementation along those lines, which as a side effect defines the following useful traits: identity<T> (::type == T) remove_cv_ref<T> (remove_cv<remove_reference>) combine_cv<T, U> (copies the cv qualifiers from U to T) common_arithmetic_type<T, U> (for arithmetic or enum types, their common type per "usual arithmetic conversions") composite_pointer_type<T, U> ( the "composite pointer type" per the standard, that is, the pointer type to which both T and U will convert) Should we add those traits to the library proper, or is it better for them to remain implementation details? The first three seem to be rediscovered on a daily basis.