Secure and Didactic Integration of LLM-Based Chatbots in Military Learning: User Story-Driven Design, Architecture, and Evaluation

This paper presents a secure and pedagogically grounded integration of Large Language Model (LLM)-based chatbot systems into high-assurance military learning environments. Building on more than 300 user stories from instructors and learners, we develop a modular, standards-based architecture that enables transparent, role-aware AI services within access-controlled Learning Management Systems. The approach emphasizes auditability, data sovereignty, and interoperability through a Common Learning Middleware using xAPI, LTI, cmi5, LOM, and Model Context Protocol (MCP). A comprehensive comparison of Retrieval-Augmented Generation (RAG) and Low-Rank Adaptation (LoRA) demonstrates that RAG reliably provides source-verifiable responses grounded in authorized training materials, outperforming LoRA across public and synthetic textbook benchmarks. LoRA offers advantages when retrieval is impractical but does not surpass RAG under multi-course, access-restricted conditions. The system further operationalizes adaptive instruction by linking learning analytics, competency mappings, and structured generation of machine-readable quizzes to provide context-aware support for learners and instructors. Overall, the work offers architectural, methodological, and empirical guidance for deploying trustworthy and extensible LLM-based assistants in regulated educational settings, with applicability beyond the military domain.