I'm not sure that's true.
At least for JPA and for BV (on the other side), in case of ambiguity, we have been able to have specific TCK tests disabled pending clearing in the next spec version. So I never experience a situation where the RI was opposed to me as the truth. Anyone has a different experience?
It's not the prettiest of examples, but we kinda did that with the JASPIC 1.1 register session feature. It was unfortunately quite a bit underspecified, so vendors basically had no clue how to implement it at all and had to follow the way it was done in the RI.
The RI is first in time by JCP definition but that's a tangential aspect.
It's also an important requirement; the spec can not be released without an implementation proving it's at least possible to implement.
This may be to prevent some of the C++ spec mayhem of a while back where it was nearly impossible to implement certain aspects in practice.