Guided Uncertainty Learning Using a Post-Hoc Evidential Meta-Model