Data information model, reproducibility and procedural accountability in the information threat monitoring system

Abstract

Analysis of recent research and publications. Scientific papers on countering disinformation distinguish approaches to documenting datasets via data statements and datasheets (E. Bender, T. Gebru), as well as models via model cards (M. Mitchell). Reproducibility in NLP systems has been studied by I. Magnusson and A. Belz, while continuous AI auditing has been addressed by J. Mökander and M. Minkkinen. However, existing publications lack an approach that integrates reproducibility, analytical assessment logging, and an audit trail directly within a unified information model of an individual message record.

The aim of the study is to substantiate a comprehensive data information model, mechanisms for ensuring computational reproducibility, and principles for building end–to–end procedural accountability in applied information threat monitoring systems.

Main body of the research. The paper proposes to consider the information model as the integration framework of the media monitoring system. An enriched news message record, structured as a four–layer object, is chosen as the modeling target. The first layer covers identification and origin fields (IDs, URLs, dates). The second layer contains primary text fields and language tags for both original and translated text. The third layer integrates the outputs of analytical modules: thematic rubrics, sentiment scores, and polarity inversion flags. The fourth layer accumulates service quality and technical support attributes (text hash, duplicate flags, versions of dictionaries, and configurations). The end–to–end data lifecycle is structured so that quality control links all stages from collection to utilization. The technical foundation for the reproducibility of results is defined as the mandatory capturing of seed values, model configurations, and linguistic resource versions in the system log. Procedural accountability is implemented via an end–to–end audit trail that records the processing chain of a message. To minimize risks when handling sensitive content, the integration of a human–in–the–loop expert architecture is substantiated, where the final decision is made and logged incorporating expert validation.

Conclusions. The study proves that the effectiveness of monitoring systems depends on the transparency and verifiability of the data architecture. The proposed information model ensures a transition from fragmented storage to a holistic integration pipeline. Capturing the technical parameters of data processing guarantees full reproducibility of results under identical conditions. The end–to–end audit trail and the expert loop provide the necessary conditions for verifying automated conclusions, distinguishing between errors, and continuously improving the system. The practical value lies in establishing a methodological framework for developing accountable information threat mitigation systems.