From Data to Evidence: AI-Enabled, Risk-Informed Decision Support for Nuclear Licensing

Drafting safety analysis reports for nuclear reactors under regulatory guidance such asNUREG-1537 requires strict traceability, consistency across heterogeneous document sets, and anauditable use of evidence. Conventional approaches based on large language models (LLMs) donot reliably satisfy these requirements; they may produce text unsupported by the applicant’s owndocumentation, introducing compliance risks in safety-critical contexts. We propose a retrieval-first,graph-augmented architecture in which the evidence chain (not the LLM) is the authoritative source.The framework couples a Chroma vector database for fine-grained semantic retrieval with a knowledgegraph capturing the document topology and model-based systems engineering architecture of the system under consideration. An orchestration layer blends these two data structures into a structured ContextPack that constrains LLM generation to content traceable to specific source documents. The role of the LLM is to synthesize and articulate evidence, not to supply it. A set of guardrails enforces citation traceability, surfaces information gaps, and flags sentences that cannot be grounded in the
retrieved evidence (making uncertainty explicit and auditable rather than eliminating it). A pilot study based on the ISU AGN-201M research reactor final safety analysis report quantitatively demonstrates the approach, in which 53 NUREG-1537 sections are drafted based on a corpus of documents that have
been collected throughout the years of operating the Idaho State University AGN-201M.