THREE-LEVEL KNOWLEDGE ARCHITECTURE AS A TOOL FOR MINIMIZING LOGICAL DRIFT IN AI-ASSISTED RESEARCH

DOI: 10.31673/2412-4338.2026.019008

Abstract

This paper addresses the problem of “logical drift” and statistical hallucinations of large language models (LLMs) in the context of fundamental scientific research. The author proposes and formalizes a method of Induced AI-Theory Expansion (IAI-TE), based on a three-level knowledge architecture: the axiomatic core (A-Core), the conceptual codex (S-Template), and the full specification. The key innovation of the method lies in transforming the generative capacity of AI from a source of error into an instrument of rigorous deduction through the implementation of artificial reality filters. A Consistency-Enforcement Protocol (CE-Protocol) is developed to ensure dual verification: textual (logical coherence) and symbolic (dimensional analysis). The practical validation of the method is demonstrated through the complete deductive reconstruction of the Temporal Theory of the Universe (TTU) from a compact 7.2 KB core. Experimental results confirm 100% successful recovery of 47 fundamental equations of the theory after 23 iterations of the CE-Protocol, demonstrating a transition from memorization to genuine deduction. The proposed method establishes the foundation for a new epistemological paradigm — AI-Resilient Science — in which scientific theories become executable algorithms capable of self-regeneration and scalable expansion without loss of logical integrity.

Keywords: IAI-TE, artificial intelligence, scientific methodology, axiomatic core, theory coherence, LLM, logical drift, AI-Resilient Science, Temporal Theory of the Universe (TTU), post-book science, algorithmic epistemology, consistency-enforcement protocol, self-regenerating theories. 

References

1. Ji Z., Lee N., Frieske R. et al. Survey of hallucination in natural language generation. ACM Computing Surveys. 2023. Vol. 55, No. 12. P. 1–38. DOI: 10.1145/3571730.
2.  Lemeshko A. Temporal Theory of the Universe – Minimal Memory Kernel (TTU_CORE_RECALL_v1.0). ResearchGate. 2024. DOI: 10.13140/RG.2.2.28830.40001.
3.  Marcus G. The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence. arXiv preprint. 2020. arXiv:2002.06177.
4.  Pearl J. Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge: Cambridge University Press, 2009. 484 p.
5.  Wolfram S. Writings: On the symbolic and linguistic capabilities of large language models. Wolfram Media. 2023. URL: https://writings.stephenwolfram.com/ (дата звернення: 12.12.2025).
6.  Wilkinson M. D. et al. The FAIR Guiding Principles for Scientific Data Management and Stewardship. Scientific Data. 2016. Vol. 3, Article 160018. DOI: 10.1038/sdata.2016.18.
7.  Krenn M. et al. On Scientific Understanding with Artificial Intelligence. Nature Reviews Physics. 2022. Vol. 4. P. 761–769.
8.  Popper K. R. The Logic of Scientific Discovery. London: Routledge, 1959. 544 p.
9.  Kuhn T. S. The Structure of Scientific Revolutions. Chicago: University of Chicago Press, 1962. 172 p.
10.  Smaragdis E. et al. AI-Driven Knowledge Discovery and Representation in Scientific Domains. AI Magazine. 2023. Vol. 44(4). P. 1–15.
11.  Chen M. et al. Evaluating Large Language Models Trained on Code. arXiv preprint. 2021. arXiv:2107.03374.
12.  Berners-Lee T., Hendler J., Lassila O. The Semantic Web. Scientific American. 2001. Vol. 284(5). P. 34–43.
13.  Stanford Institute for Human-Centered Artificial Intelligence. AI Index Report 2024. Stanford University, 2024. URL: https://aiindex.stanford.edu/report/
14.  Vaswani A. et al. Attention Is All You Need. Advances in Neural Information Processing Systems. 2017. Vol. 30. P. 5998–6008.