Under the paper's strictest target-concept-disjoint control, the held-out test sets are small: only $21$ unique held-out targets on LectureBank-Full and $18$ on MOOC-CS, and these targets all come from the same two question families. As a consequence, even strict-parity headline numbers obtained under this control support only a narrow claim about curated, template-based prerequisite QA rather than broad graph-retrieval generalization.