Researchers presented the GVD framework for document version management

Researchers have published GVD (Governed Versioning and Deduplication) — a framework for automated versioning and deduplication of corporate documents. The system allows tracking changes in rules and policies, identifying inconsistencies and duplicates without using LLM. The system's operation is based on the Counterfactual Span Probing (CSP) method, which refines the classification of relationships between documents. During testing on a set of 120 corporate documents, divided into 59 version families, GVD demonstrated an F1-measure of 0.97 when building version families and 0.94 when checking rule consistency. The use of CSP allowed increasing the consistency accuracy from 0.90 to 0.94. The GVD architecture works completely locally, eliminating the need for cloud computing or third-party language models. The system assigns incoming documents to corresponding families through bidirectional rule matching, comparing them with the "family memory" to identify clarifications, inconsistencies, or new information. The framework is oriented towards auditing: it keeps a history of changes as an audit trail, suppressing duplicates and passing only significant updates for manual verification. The work is published under the CC BY 4.0 license, which allows using the proposed algorithms to create knowledge management systems in organizations with high document turnover.