| Educational organizations adhering to management standards such as ISO 21001 face significant challenges in producing compliant documentation that satisfies both pedagogical relevance and strict regulatory constraints. While Large Language Models (LLMs) excel at content generation, they often struggle with structural rigidity and factuality—a phenomenon known as hallucinations—rendering them unreliable for certification audits. In this paper, we propose a Neuro-Symbolic Retrieval-Augmented Generation (RAG) architecture designed to automate the production of ISO 21001-compliant documents (e.g., syllabi, audit evidence, improvement plans). Our approach bridges the gap between connectionist AI and symbolic logic by introducing a Hybrid Knowledge Base formalized as a set of Symbolic-Vector Objects. We define a constrained decoding mechanism where an ontology enforces regulatory guardrails on the LLM output. Through experiments on a synthetic dataset validated by certified auditors, our system achieves a 98% Structural Compliance Score and reduces hallucinations to 1.5%, significantly outperforming standard RAG baselines. We present the formal model and the system architecture, demonstrating how explicit symbolic constraints can govern generative processes to ensure audit-readiness. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.