Opaque timetabling heuristics often deliver strong schedules yet remain difficult to interpret, limiting trust, diagnosis, and controlled adaptation in educational scenarios. This work presents a practical pipeline that learns a constraint-aware graph neural network (GNN) surrogate from input-output pairs of a black-box timetabling solver and then applies global explainability to characterize the surrogate’s decision-making logic across entities and relations in the context of timetabling. Using a synthetic dataset with hard constraints, the study evaluates both the assignment fidelity to the solver outputs and the feasibility under hard constraints, complementing these with global explanations and counterfactual sensitivity analysis. The results highlight which entities and relations most influence the predicted assignments. The pipeline is intended as a methods-oriented contribution that standardizes data generation, surrogate training, and explanation estimators for timetabling, enabling reproducible assessment of interpretability alongside fidelity and feasibility without claiming domain deployment readiness.