CHERI: An Agentic AI Framework for Structural Quality Control of Excel-Based Budget Impact Models

Objectives: Excel-based health economic (HE) models encode their logic implicitly across worksheets, formulas and layout, which limits what large language model (LLM) agents can reliably understand and check. We developed CHERI, an agentic, neuro-symbolic framework that builds an ontology-grounded knowledge graph (KG) of Excel-based budget impact models (BIMs), and evaluated it for KG construction and structural quality control (QC). Methods: We distinguish technical, structural and conceptual QC, and define structural QC as verifying that a workbook implements its model class correctly, both computationally and presentationally. CHERI parses a workbook, uses a neuro-symbolic loop in which an LLM proposes how its calculations correspond to the computational spines of a BIM, accepts only proposals that pass a deterministic proof against the workbook’s own arithmetic, and assigns ontology terms to the verified structure. We evaluated CHERI on HELMET-BIM, a benchmark of four de-identified industry BIMs, each in three versions seeded with computational and formatting defects (12 workbooks). Results: CHERI fully bound 10 of 12 workbooks, with spine-level precision of 99.3% and recall of 98.7%, and explained at least 97.7% of the formula cells feeding the budget impact in every workbook. QC agents reasoning over the KG detected 36 of 42 seeded computational defects (sensitivity 85.7%; 97.0% for major defects) and 15 of 18 seeded formatting defects, and identified 12 defects already present in the original workbooks. Conclusions: A verified, persistent representation of model structure allows LLM agents to perform structural QC of Excel-based BIMs that cell-by-cell review cannot. The approach provides a foundation for agents that check, adapt and translate HE models across model classes. Keywords: budget impact analysis; quality control; model verification; large language models; agentic AI; knowledge graph; neuro-symbolic AI; Excel; health economic modeling.