Not using it as a post processing scheme is IMO the proper way to utilize meta model feature exposures, just as a metric. The problem here is when you incorporate data from the past, in this case, meta model feature exposures into your model in an effort to try extract uniqueness of alpha vis a vis the meta model, you are introducing an additional constraint in your optimization process that limits your search parameters to focus on neutralizing feature exposures which muddles the natural churn of your model thus often leads to overfitting. You are also assuming meta model feature exposures will be persistant in the near future, it may or may not.