Authors - Miroslav Stefanov, Stoyan Denchev, Kristiyan Stefanov Abstract - Large Language Models (LLMs) are increasingly used in educational settings, but they are not inherently constrained to the boundaries of specific instructional materials. This can lead to unsupported claims, external knowledge leakage, and reduced instructional precision. This study proposes and evaluates a lesson-bounded LLM architecture for reliable educational AI systems. The architecture combines retrieval-augmented generation, context restriction, structured response control, explicit refusal behavior, and post-hoc confidence calibration. Using a multi-domain instructional dataset and a benchmark of inscope and out-of-scope questions, the proposed system is compared against an unconstrained baseline LLM. Results show strong retrieval discrimination and boundary control, with high Area Under the Receiver Operating Characteristic Curve, high Average Precision, strong refusal recall, low out-of-scope answer rate, reduced verbosity, and improved support-based instructional density. Calibration analysis further shows that raw retrieval scores are not reliable probability estimates, but Platt scaling substantially improves confidence reliability. These findings suggest that lesson-bounded architectural constraints can improve the controllability, auditability, and reliability of intelligent educational systems while highlighting the need for stronger factuality evaluation and confidence interpretation.
Open Zoom