The paper considers the task of automated verification of Russian national project activities by their quantitative outcomes. The source of information is a package of reporting documents submitted by the executor to the oversight body. The composition of the package is not known in advance, and the documents are heterogeneous in type and informational nature. A simple summation of numerical values from such documents yields an inflated result, since one transaction is typically reflected in several documents at once. The paper proposes a verification architecture that uses three sequential paths: summation over a series of uniform primary documents, comparison with the maximum among aggregate documents, and a deduplicated sum obtained with a language model. The paths are ordered by decreasing degree of trust in the result. The architecture has been implemented within a software pipeline and tested on real document packages. The scientific novelty lies in the hierarchical organization of paths, where the probabilistic method is engaged only when the other methods have failed.
automated audit, national projects, large language models, quantitative verification, transaction deduplication, on-premise large language model
1. LayoutLM: Pre-training of Text and Layout for Document Image Understanding / Y. Xu [et al.] // Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020. P. 1192–1200.
2. OCR-Free Document Understanding Transformer / G. Kim [et al.] // Proceedings of the European Conference on Computer Vision. 2022. P. 498–517.
3. Language Models are Few-Shot Learners / T.B. Brown [et al.] // Advances in Neural Information Processing Systems. 2020. Vol. 33. P. 1877–1901.
4. Llama 2: Open Foundation and Fine-Tuned Chat Models / H. Touvron [et al.] // arXiv preprint arXiv:2307.09288. 2023.
5. Dudihin V.V., Kondrashov P.E. Metodologiya ispol'zovaniya bol'shih yazykovyh modelej dlya resheniya zadach gosudarstvennogo i municipal'nogo upravleniya po intellektual'nomu referirovaniyu i avtomaticheskomu formirovaniyu tekstovogo kontenta // Gosudarstvennoe upravlenie. Elektronnyj vestnik. 2024. № 105. S. 5–18.
6. Natarova E.V. Cifrovaya transformaciya v sisteme gosudarstvennogo audita // Nauchnyj vestnik: finansy, banki, investicii. 2024. № 4 (69). S. 76–88.
7. Christen P. Data Matching: Concepts and Techniques for Record Linkage, Entity Resolution and Duplicate Detection. Heidelberg: Springer, 2012. 270 p. DOI:https://doi.org/10.1007/978-3-642-31164-2
8. Köpcke H., Rahm E. Frameworks for entity matching: A comparison // Data & Knowledge Engineering. 2010. Vol. 69. № 2. P. 197–210. DOI:https://doi.org/10.1016/j.datak.2009.10.003
9. Limanova N.I., Seleznev I.A. Analiz effektivnosti klient-servernoj arhitektury // Byulleten' nauki i praktiki. 2022. T. 8. № 7. S. 392–396. DOI:https://doi.org/10.33619/2414-2948/80/37
10. vLLM. Official documentation. URL: https://docs.vllm.ai/en/latest/ (data obrashcheniya: 21.05.2026).
11. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models / J. Wei [et al.] // Advances in Neural Information Processing Systems. 2022. Vol. 35. P. 24824–24837.
12. Self-Consistency Improves Chain of Thought Reasoning in Language Models / X. Wang [et al.] // Proceedings of the International Conference on Learning Representations. 2023.
13. Efficient Memory Management for Large Language Model Serving with PagedAttention / W. Kwon [et al.] // Proceedings of the 29th Symposium on Operating Systems Principles. 2023. P. 611–626. DOI:https://doi.org/10.1145/3600006.3613165
14. Zadornov A.A., Nikonov S.S., Fedotov M.V. Vliyanie edinogo informacionnogo prostranstva na avtomatizaciyu processa polucheniya i obrabotki informacii // Vestnik ekonomicheskih i sociologicheskih issledovanij. 2024. № 1. S. 21–27.
