Regulatory evidentiary closure as a formal framework for citation faithfulness in large language models
Large language models are increasingly used to generate answers with citations in regulatory settings, including compliance review, clinical governance, fiscal assessment, legal interpretation, and institutional risk analysis. Existing work on attribution, retrieval augmented generation, citation evaluation, and sufficient context has clarified when generated claims are grounded in identifiable sources. Regulatory use raises a stricter question. A citation may support an atomic sentence while still failing to justify the regulated conclusion that the system invites a user to accept. This paper develops regulatory evidentiary closure, a conclusion-relative standard for that question, as a formal framework. The framework defines source existence, domain-indexed claim support, and closure as relations over an authorized corpus, cited source set, generated claims, material elements, and regulated conclusion, and states them in explicit notation. Three propositions record the framework’s structure: claim support does not imply closure, metrics restricted to internal model signals cannot determine closure without domain rule structure, and closure checking reduces to a finite coverage problem over material elements. Worked examples from HIPAA security compliance and clinical guideline recommendation show the same failure structure across domains, where a citation supports an atomic claim while the conclusion’s material elements remain uncovered. A coverage matrix then maps seventeen existing evaluation methods against five evidentiary components and shows that none states conclusion-relative closure as an evaluation object. The paper translates the framework into a nine-step reviewer-facing audit protocol and identifies future work on element-mapping reliability, benchmark design, and tool-assisted closure assessment.
Authors
- Harish Gupta (ORCID: https://orcid.org/0000-0002-1390-9809)
Institutions
- The University of Texas at Austin (US)
Publication Details
- Journal
- Discover Artificial Intelligence
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1007/s44163-026-02305-w
- Primary Topic
- Topic Modeling
- Type
- article
- Field-Weighted Citation Impact
- 0.00