Local Coherence and Global Drift in Long-Term Research with Large Language Models: Persistent-Dependency Violations and a Control Architecture for Research-Process Integrity

Abstract

This study examines global drift of the research state in long-term human research work with large language models (LLMs): a situation in which individual steps remain locally admissible and internally coherent, yet their sequence leads to the loss or alteration of previously established foundations of the research. The analytical object is constituted by Active Persistent Dependencies—relations among objects, statuses, sources, verification results, conclusions, and decisions about subsequent transitions that must remain in force across multiple stages of the work.The empirical basis was a documented internal corpus of long-term research projects comprising eight base cases, eight negative-control cases, and additional materials. The base cases were classified using six pre-established criteria; one positive case additionally underwent an independent verification of classification reproducibility. Seven of the eight base cases were assigned to the class under consideration, one was classified as borderline, and all negative-control cases remained outside it.The analysis showed that Global Drift can arise at different levels of the research process: through changes to an object or its status, transfer of source foundations, interpretation of a verification result, transition from a result to a claim, description of a method's role, development of an argument, and continuation of a research stage. On this basis, seven levels of Persistent-Dependency violation and five principal control functions were identified: preservation of the Research State, verification of Active Persistent Dependencies, assessment of the meaning of verification results, decision-making on material changes, and limiting the propagation of a detected violation.The principal conclusion is that the integrity of long-term research involving LLMs depends not only on the correctness of individual steps but also on preservation of the Persistent Dependencies among them. The proposed scheme links the locus of a violation to the required control function without prescribing a single mandatory set of tools for all projects. The findings apply to the studied internal corpus; their Transferability and the effectiveness of the control scheme require further prospective testing in independent projects, other subject domains, and work with other models and technical environments.