Online learning platforms contain large question banks, yet learners often lack clear guidance on what to practice next and why a particular item is appropriate. Existing educational recommender systems can optimize learning paths, but their decisions are often difficult for learners to interpret. Large language models (LLMs) can generate natural-language explanations, but unconstrained generation may introduce hallucinated learner histories or unsupported pedagogical claims, which is problematic in educational scenarios. To address this issue, we propose an evidence-grounded closed-loop educational recommendation framework that combines lightweight learner-state tracking with controllable generative explanations. In this system, a deterministic ranking module selects candidate exercises using pedagogical signals such as mastery gaps and learning efficiency. The LLM is restricted to retrieved learner logs and instructional metadata and must provide explanations with explicit evidence references. We evaluate the framework through system benchmarks, simulation-based ranking comparisons, and manual assessment of explanation quality. The results suggest that the framework is feasible for online deployment, that the hierarchical efficiency strategy improves skill coverage and learner differentiation in simulation, and that evidence-grounded prompts improve explanation usefulness while reducing observed fabricated-history cases in sampled outputs. Overall, the study shows that separating deterministic recommendation from evidence-supported explanation can improve the credibility and practical utility of LLM-assisted educational recommendation.