The canonical prerequisite QA splits used by this paper are heavily templated and share substantial surface forms across train and test. On LectureBank-Full the canonical split shares $92$ exact train–test questions and $80$ train–test target concepts. On MOOC-CS it shares $68$ exact questions and $60$ target concepts. These counts motivate reporting question-disjoint and target-concept-disjoint controls rather than relying on the canonical splits, because a system can otherwise match memorized surface forms.