Operating procedures are essential for safe control room operations in nuclear power plants. However, many legacy procedures remain in paper-based formats, limiting their integration with digital control systems, automated operation, and operator decision-support systems. Transforming these documents into machine-interpretable structures is therefore necessary, yet challenging due to complex layouts, deeply nested procedural hierarchies, and heterogeneous formatting patterns.

This paper presents an LLM-assisted framework for reconstructing structured representations of nuclear power plant operating procedures. The framework defines explicit structural constraints that specify the output format and hierarchical relationships, thereby guiding the LLM to generate consistent machine-interpretable representations. To further improve reconstruction reliability, the framework incorporates systematic comparative assessment against reference procedures, enabling human-in-the-loop review and iterative refinement of reconstruction rules and outputs. A case study is conducted to evaluate the proposed approach through both qualitative and quantitative analyses. The evaluation investigates reconstruction errors and assesses structural reconstruction performance by comparing the generated outputs with human-curated references.

This work provides practical insights into the application of LLMs for structured procedural digitization and highlights a viable pathway toward reliable LLM-assisted document reconstruction in safety-critical domains.