Two leakage-controlled splits for prerequisite-QA evaluation, stated as plain definitions at first use:

• A question-disjoint split removes any test instance whose question string (or near-duplicate) appears in training. This blocks gains that come from memorizing question text rather than reasoning over the prerequisite graph.
• A target-concept-disjoint split is strictly stronger: in addition to removing question-string overlap, it removes test items whose gold target concept IDs appear as training supervision. This blocks gains that come from having seen the held-out target concept during training.

The two splits are the concrete instruments by which the leakage audit's two overlap kinds (question-string and target-concept) are operationalized; they are used together in headline retrieval comparisons so that surviving gains can be attributed to the retrieval policy rather than to overlap with training material.