According to a spokesperson for DOE’s Office of Nuclear Energy, subject matter experts who reviewed the document said the results were “consistent with what would be expected for a Revision 0 document.”
A pilot program to use Documented Safety Analysis (DSA) documents approved by willing participants is planned as a future step, which will bring this capability closer to production-grade regulatory use. This requires extending the work into a secure operating environment compared to a proof-of-concept project using open source documentation.
This project was a collaboration between DOE, Everstar, the State of Idaho, Argonne National Laboratory, and Microsoft.
AI acceleration: The project emerges in the context of the Genesis mission, which was launched in November 2025 as a national effort to accelerate the use of AI in scientific research.
A DOE-NE spokesperson said the agency will use AI tools to prepare the draft and shift its expert workforce to refine and review it.
This document was prepared for the General Purpose High Temperature Gas Reactor (HTGR) at the National Nuclear Reactor Innovation Center at INL. The team used an INL site-level DSA, a generic HTGR preliminary DSA (PDSA), and a demonstration of the safety analysis of a microreactor experiment (a document written by a subject matter expert) as inputs to the AI. We also had access to several standard reference corpora, including the NRC agency-wide document access and management system database and publicly available federal regulations.
Everstar’s AI tool, named Gordian, uses a search-augmented generative architecture that generative models use to produce structured and cited output for NRC compliance. This approach reduces illusions by anchoring the response to the source document, which serves as the technical foundation.
A DOE-NE spokesperson said experts have also created “skills modules.” It is a structured set of instructions that governs the behavior of the AI at each stage of the document generation process, including enforcing citation-only disciplines, defining regulatory mappings, specifying requirements by section, self-checking and quality assurance routines, and more.
“As the PDSA is an inherently preliminary and incomplete document, the completeness of the AI output is limited accordingly and is a direct reflection of the input and not a failure of the tool,” a DOE-NE spokesperson said.
The spokesperson also said Gordian’s ability to identify missing, derived, or inconsistent information across source documents is a capability advantage, surfacing issues that might otherwise persist through multiple human review cycles.
They said the tool shows room for improvement in accurately distinguishing between a parameter’s safety limits and its operating range, which is of technical and regulatory importance in nuclear licensing.
What’s next: Several next steps have been identified by DOE-NE.
An immediate priority is to develop dedicated review tools to systematically document and resolve discrepancies.
The team is in the process of launching a formal NRC working group on AI-assisted licensing.
Documentation standards may evolve with AI tools to ensure that source documents are written in a way that AI can properly parse, source, and cite.
In the long term, the same modular architecture is envisioned for additional final safety analysis report chapters, PDSA generation, NQA-1 (Nuclear Quality Assurance) compliance documentation, and fuel manufacturing facility licensing.
