The current transition to so-called "data-driven management," based on the creation of full-scale digital twins of complex socio-economic systems, faces a fundamental challenge. Managed systems are ergatic, meaning they include ultimately non-formalizable human, technological, and environmental components. This article demonstrates that optimization problems solved for such digital models are inherently ill-posed with respect to the original, as the human element introduces reflexive non-formalizability and implicit knowledge that cannot be captured by any formal structure. Risk in this paradigm ceases to be an external threat, becoming an immanent property of the management system itself, generated by the inevitable and growing discrepancy between the trajectories of the virtual model and the real ergatic process. Strategies optimal for the "Digital Leviathan" (total simulation) lead to the destabilization of the original, erosion of the human component, and the system exceeding the boundaries of controllability. The authors substantiate the need for a paradigm shift: from searching for a point optimum for an incomplete model to synthesizing robust and adaptive strategies that explicitly minimize the risk functional. A key element of the new "symbiotic management" paradigm is the formal inclusion of the human operator as a corrective loop in the iterative dialogue between the model and reality. The proposed architecture of symbiotic management, grounded in Bayesian updating and minimax principles, enables not merely risk minimization but active management of epistemic uncertainty, transforming human intuition into a structured source of model correction. This approach ensures the resilience of complex socio-technical systems under conditions of fundamental knowledge incompleteness, offering a path from the illusion of absolute control to the wisdom of collaborative navigation. The goal of management becomes not an absolute optimum in virtual space, but dynamic adaptive efficiency in the real world, achieved through constant management of the immanent risk of trajectory divergence. The article introduces and contrasts two concepts: the "Digital Leviathan" (the temptation to manage the simulation) and the "formalism of risk" (the discipline of constantly comparing the map and the territory).

This study examines the primary limitations of tree-based decision models in the insurance sector, namely their sensitivity to noise and outliers, susceptibility to overfitting, and limited generalization capabilities. To overcome these challenges, we propose leveraging advanced machine learning techniques, neural networks, and stacking ensemble methods. We hypothesize that the quality of loss prediction can be slightly improved by combining the predictions of multiple base models (XGBoost, CatBoost, Logistic Regression, Random Forest, and TabTransformer), alongside the original dataset features, into a more robust ensemble. Specifically, the meta-model is constructed as a fully connected neural network with an additional input layer for the original features. The empirical study utilizes Russian auto insurance data from 2023 to 2025, detailing the experimental framework for training heterogeneous machine learning and neural network models. The data preprocessing pipeline is described, with a particular focus on a specific categorical encoding scheme based on the European car classification by luxury segment. Special attention is given to the severe imbalance of the target class, as losses account for less than 3% of the total dataset. Based on the trained base models, ensemble models are constructed using five stacking techniques: Simple Average, Constrained Logistic Regression, Adaptive Regression by Mixing (ARM), ARM-Tweedie, and Multilayer Perceptron (MLP). A comprehensive comparative analysis of the base models and the ensembles is conducted using both classification metrics (ROC-AUC, PR-AUC) and probabilistic metrics (Log Loss, Brier Score). The results indicate that the best-performing base model is TabTransformer (Brier Score = 0.037, Log Loss = 0.185, ROC-AUC = 0.635, PR-AUC = 0.054). The MLP-based ensemble outperformed it across all key metrics (Brier Score = 0.026, Log Loss = 0.131, ROC-AUC = 0.649, PR-AUC = 0.055). These findings confirm the initial hypothesis: the fully connected neural network ensemble yields a slightly improvement in the primary evaluation metrics. This demonstrates the methodological value of the proposed approach: even under conditions of a weak data signal, the neural network ensemble, augmented with original features, ensures a more effective aggregation of predictions compared to traditional predictive analytics and alternative stacking ensemble methods.

The relevance of this research stems from the growing need to automate corporate meeting minutes, a labor-intensive and time-consuming process with a high risk of errors. Current speech recognition systems (ASR – Automatic Speech Recognition) perform poorly in multi-user conversations, with overlapping voices, background noise, and the use of specialized vocabulary. Speaker identification (SID – Speaker Identification) is particularly challenging when there are few reference recordings per employee, which is typical in an office environment. Acoustic environments also pose additional challenges: echoing offices, simultaneous speech, and poor voice separation. Therefore, the goal of this study is to develop a hybrid neural network model for accurate transcription and reliable speech attribution in group discussions. This paper analyzes the weaknesses of ASR systems and SID methods and proposes a three-tier architecture: audio stream segmentation (VAD – Voice Activity Detection, SCD – Speaker Change Detection), voice embedding identification with a reference attention mechanism, and speaker-aware transcription. A mathematical justification for the key components is also provided. The theoretical significance lies in the development of an attention-based SID method that effectively solves few-shot learning problems. Unlike traditional approaches that require large labeled corpora, the new architecture learns to match speech fragments with a small number of reference recordings. The effectiveness of the method has been empirically confirmed using real corporate meeting recordings (over 20 hours, five participants). The segmentation module demonstrated a VAD F1 score of 0.885, and the speaker identification module achieved an accuracy of 62.5%, which is three times higher than chance (20%). Transcription was performed using the external Whisper system, which guarantees high recognition accuracy. The results enable the development of a system for automatic meeting minutes, and the architecture supports a large number of participants and is easily adapted to various communication formats: from negotiations to client meetings. The functionality of the hybrid neural network model can be used to create various tools for corporate meeting minutes 

The problem of a systematic and scientifically-based approach to the organization of e-learning is becoming increasingly relevant in the context of the widespread use of large language models. This article explores the issue of conceptual and methodological support for the operation of Intelligent Tutoring Systems (ITS). It highlights the challenge of describing and explaining the internal mechanisms of these systems in scientific publications, allowing for the verification and automatic explanation of their decisions. As an example, the author's experimental learning system, AESU, is examined. The article focuses on the fundamental concepts, mechanisms, and methodological solutions employed in this system. The main concepts include P.K. Anokhin's theory of functional systems (the model of afferent synthesis), D.A. Pospelov's applied semiotics and V.A. Lefebvre's theory of reflexive control. The role of cognitive visualization (mapping) in the synthesis of solutions is demonstrated. This involves combining the logic of production expert systems, fuzzy logic, and Shortliffe's hypothesis testing, which are the intellectual mechanisms of the learning system's solver. The emergent effect arising from the use of the above-mentioned concepts and mechanisms allows us to propose an author's approach to the implementation of modular data concentration and cross-cutting approach of the learning situation. The structural and functional notation of maps becomes the system-forming element in the synthesis of ITS solutions: the method of Cognitive Maps of Knowledge Diagnosis and the method of Unified Graphic Visualization of Activity. The article demonstrates the possibility of mutual transition between the parametric and visual representation of the learning situation between these types of maps and their use in the development of solutions by both the solver of the learning system and the student during their work in the electronic learning environment. In conclusion, the architectural features of the AESU ITS are noted. In addition to the typical functions of learning management systems, the proposed ITS implements a dialogic mechanism for interacting with students in a natural language form without using the capabilities of large language models. The flexible feedback mechanism allows for interactive dialogue based on methodologically sound conclusions, which increases the level of trust between the student and the learning system and promotes engagement in the educational process based on the ITS.

This paper presents a simulation of the spatial redistribution of optical field intensity in a planar waveguide structure formed by the interface between a nonlinear medium and a medium with a graded refractive index. A generalized dependence of permittivity on the distance from the contact surface of the media was chosen as a model for the gradient medium. A medium with a linear optical effect (Pockels effect), in which permittivity is linearly dependent on the amplitude of the electric field, was chosen as a model for the nonlinear medium. The study identifies critical operating modes in which the maximum intensity of the guided mode can be deliberately located within any of the constituent media or precisely at their interface by adjusting the optical parameters of the system that determine the critical value of the effective refractive index. The main result of the study is proof of the existence of a critical effective refractive index regardless of the choice of nonlinearity model, at which the field maximum is precisely localized at the interface. It is shown that the value depends solely on the parameters determining the profile of the graded refractive index, which ensures the dispersion relation. Particular emphasis is placed on identifying how each of the gradient profile parameters individually affects the magnitude of the transition threshold. Аanalysis revealed that the critical value of the effective refractive index varies monotonically with changes in the characteristic spatial scale and dielectric parameters of the graded layer. It is shown that the nature of this influence is determined by the sign and magnitude of the corresponding dielectric constants, allowing for the targeted formation of the desired dependence of the critical refractive index on the optical parameters of a planar waveguide structure. Unlike most studies, which are limited to numerical modeling of specific configurations, this study proposes a rigorous analytical approach that allows for the derivation of explicit criteria for the existence of a critical regime. The potential use of the obtained results for calibrating experimental setups, adjusting phase-matching conditions for second-harmonic generation, and estimating acceptable parameter tolerances in integrated circuit manufacturing is also discussed. The results provide a theoretical basis for the development of methods for controlled light localization in hybrid waveguide structures by varying their material and geometric parameters or the operating wavelength.

With the constantly growing number of online platforms and intensifying competition among marketplaces, optimizing their operations and improving management efficiency are becoming crucial tasks. Mathematical modelling of sales processes on marketplaces allows predicting changes in conditions and minimizing risks. Models must consider various factors such as demand seasonality, competition, changes in logistics, and marketplace policies. This helps suppliers optimize prices, inventory, and logistics processes, reducing financial losses and enhancing business stability. Marketplaces like Amazon, Wildberries, or Ozon process an enormous number of transactions and customer interactions daily, requiring the development of mathematical models for management systems and forecasting. This article explores mathematical models based on queuing systems (QS) describing their principles and potential applications in e-commerce. For predicting customer flows and optimizing the sales funnel on marketplaces, single- and two-phase queuing models with repeated requests are proposed, allowing forecasting parameters of planned marketing campaigns and assessing their impact on revenue and customer base. For modelling the sales funnel, multi-phase models are suggested, estimating the time a potential buyer spends at each funnel level, enabling evaluation of the effectiveness of different funnel stages and identifying bottlenecks that can reduce conversion; developing recommendations for sales funnel optimization, such as changing interface design, improving search systems, or offering more favourable payment terms; and predicting sales volume and optimizing product inventory based on analysing application flows at different funnel stages. The article lays the foundation for further research on adapting these models to real processes, including empirical validation, parameter determination, and algorithm development for optimizing seller sales strategies.

The aim of this study is to develop an optimization model for determining the rational livestock structure and population size for various livestock species in a given area, taking into account farm specialization and the environmental load on land resources. This is the first model developed that allows for determining the livestock structure and population size under conditions of pasture livestock farming, both with and without pasture land degradation, which is relevant for Mongolian agriculture. The model's objective function is profit-oriented. Constraints include production volumes, provision of animals with feed and nutrients, labor costs for animal maintenance, water supply, environmental requirements for pasture use by animals, and the population size of different groups. Since the supply of feed for livestock is determined by the yield of forage crops on pastures, the dynamics of forage crop yields on pastures in different regions of Mongolia are analyzed based on data from 1960 to 2024. It is shown that harvest volumes from 1992 to the present can be described using an exponential function for the Western, Khangai, Central, and Eastern regions of Mongolia. The model was implemented using households with different industry specializations. The following livestock species were analyzed: sheep, goats, horses, cattle, and camels. A certain average farm size was identified based on pasture area and labor costs for production. The production, economic, and environmental characteristics of the model were determined. Optimal solutions were calculated for households with different specializations, with and without pasture degradation. A comparative analysis of the results was conducted, and the best options for maximum profit were obtained. The LP Solve application was used to implement the developed model. The developed model is useful for production planning with different livestock specializations and varying levels of pasture degradation.

The problems of fault estimation in dynamics and sensors of technical systems described by linear dynamic models with time invariant parameters described by differential equations subject to the unknown external disturbances and measurement noise in sensors are studied. The purpose is to design a device using information about inputs and readings to estimate values of faults both in dynamics and sensors. Such information is necessary to compensate the influence of these faults by fault tolerant control systems based on control correction. A solution is based on so-called interval observers of minimal dimension insensitive to the external disturbances. Such observers are constructed based on the reduced-order model of the original system insensitive to the disturbances as well. To construct the reduced-order model, the appropriate relations are used based on the Jordan canonical form describing dynamics of the model and observer. The Jordan canonical form is necessary to use the appropriate properties of the observer. Based on such observer and differentiator which produces exact estimate in finite time, the interval estimate relations are derived producing the relations to estimate faults both in dynamics and sensors. Novelty of the paper is that unlike typical way of using the interval observer, a device for estimation is constructed based on the reduced-order model insensitive to the external disturbances. This allows reducing the computational complexity of the estimation procedure and increasing the accuracy of estimation due to insensitivity to the disturbances. Theoretical results are illustrated by an example of fault estimation in the electrical actuator of manipulation robot. Simulation based on the package Matlab confirms a correctness of theoretical results. It is shown that using the reduced-order model allows reducing the computational complexity of the estimation procedure. The obtained relations can be used to design fault tolerant systems.

Optimal distribution of electric loads between energy sources is a crucial process enabling the minimization of financial costs for the energy supplier at the current demand level. This process, internationally known as economic dispatch, is traditionally built on the principle of equality of incremental fuel costs, which allows for determining the optimal power allocation in terms of achieving economic efficiency. However, the modern state of the energy sector, characterized by global ESG transformation and the implementation of Corporate Social Responsibility (CSR) principles, requires considering not only direct economic indicators but also socio-environmental ones. The latter are often hard to formalize or entirely non-formalizable, introducing significant complexity to the optimization process. These conditions necessitate the involvement of a Decision Maker (DM) with their inherent internal value system. Under these circumstances, the traditional cost-minimization approach becomes insufficient, dictating the need for new hybrid methods. This article proposes a hybrid method for power allocation based on the synthesis of the classical Lagrange multiplier method, generalized by Karush–Kuhn–Tucker (KKT) conditions, and Multi-Attribute Utility Theory (MAUT). The scientific novelty of the approach lies in the direct integration of a multi-attribute value function, reflecting the DM's subjective preferences, into the structure of the optimization problem. To validate the developed hybrid method, an isolated power system with a total load of 500 MW was modeled. The optimization problem in a static formulation was solved for two scenarios: a baseline scenario (minimizing fuel costs) and a socially-oriented scenario (considering DM participation with an active constraint on the value function level V ≥ 0.60). The results showed that meeting the DM requirements at a value level of 0.60 leads to a substantial reconfiguration of the energy balance, substituting 43% of base-load coal generation with gas power capacity. It is demonstrated that the resulting cost increase (by 22%) is compensated by achieving the required level of social value. The calculated values of dual variables (Lagrange multipliers) are interpreted as "shadow prices" of non-formalizable constraints, reflecting the marginal cost of improving the socio-environmental situation.

The paper addresses the problem of integrating heterogeneous mathematical models used in the formation of a digital model of a digital twin of a wind power plant. The relevance of the study is determined by the fact that a wind power plant is not a single computational object, but a complex technical system comprising a number of interrelated subsystems: a wind turbine, an aerodynamic circuit, a mechanical transmission, power electronics, an energy transmission system, a diesel power plant, battery energy storage systems, and consumers. For each of these subsystems, the scientific literature presents various classes of models that differ in terms of input data, level of detail, computational time step, adopted assumptions, constraints, and scope of applicability. In the development of a digital twin, such models should not only be implemented within a single software system, but also coordinated with one another in terms of parameter semantics, units of measurement, temporal characteristics, boundary conditions, and rules for transferring computational results. The aim of the study is to substantiate the feasibility of applying an ontological approach to model integration within the digital model of a digital twin of a wind power plant. It is shown that, in the absence of explicit formalization of the subject area, the increasing complexity of a digital twin is accompanied by data duplication, the emergence of implicit dependencies between computational modules, difficulties in maintaining the software system, and growing costs associated with the development of new functionality. An ontological model is proposed as a means of limiting these effects, defining a common layer of terms, relationships, constraints, and data requirements. It is advisable to apply such a model at the stage of designing new digital twin functionality, including the implementation of new computational models, modification of visualization components, database expansion, and refinement of relationships between subsystems. The novelty of the proposed approach lies in considering ontology not only as a means of representing knowledge about the subject area, but also as a tool for reducing organizational and software engineering risks in the integration of digital twin models. Particular attention is paid to the fact that an ontological model makes it possible to capture specialists’ knowledge in a formalized form and thereby reduce the project’s dependence on a limited number of employees who simultaneously possess expertise in software engineering, mathematical modeling, and the specific features of energy facilities. The paper substantiates that the use of an ontological approach can contribute to improving the consistency of digital model components, reducing the probability of errors in data exchange between computational modules, and simplifying the further development of the digital twin of a wind power plant.

The dynamic transformation of human-machine interaction constantly requires the development of new decision-making models to maintain the efficiency and quality of any information process. Public institutions, as the owners of the budget execution process, urgently need a universal methodology for assessing the performance of their accounting information systems (AIIS) in the face of dynamic IT development and high management turnover. To address this challenge, a systems approach was used to comprehensively analyze the processes associated with information processing in centralized accounting departments (CADs) and the management processes of such institutions under budget constraints. The activities of AIIS users are determined by a number of indicators grouped into four areas: subject, qualitative, budgetary, and economic. As a result of the study, a general mathematical model and target function for the functioning of the AIIS, based on the principles of the transport problem, were described and schematically presented. This model takes into account the key functions, dependencies, and indicators of the information process at the user level. A schematic model for calculating the objective function, presented in IDEF0 format, includes processes for quantifying user-processed data, calculating unproductive time losses, and total unproductive costs. The most important indicators directly impacting the productivity of the ISBU user are the time costs associated with identifying their own errors and re-entering (correcting the original information). The selection of the target criterion of the central bank determines not only operational management objectives but also a strategy for improving the efficiency of budget expenditures, allowing budget funds to be directed toward those purposes that improve the objective functions of the ISBU during its operation. The target criterion of the public entity, as the founder of the central bank, determines the budget constraint for the objective function of the ISBU, conditioned by the reduction of any expenses on supporting functions accompanying the core business. The model proposed in the study can be further used to analyze and classify users based on their professionalism, and to develop algorithms and intelligent decision- support tools at various management levels.

The article presents the methodological outcome of the first stage of a research program aimed at improving inventory and working capital management for sellers operating on electronic marketplaces. The relevance of the study stems from the fact that practical tasks of allocating limited working capital across numerous product items are solved under conditions of stochastic demand, pronounced seasonality, heterogeneous volatility and constrained computational resources, whereas traditional inventory management methods either remain static or require a priori specification of demand and cost parameters. The paper proposes a comprehensive methodology for demand analysis and working capital allocation that comprises three related components: intelligent demand analysis based on seasonal-trend decomposition and machine learning models for time series (DLinear, NLinear, XGBoost); extended analysis of demand volatility using two variants of the coefficient of variation and adaptive clustering of the assortment matrix; and a working capital allocation algorithm built on an integral «profitability index» and supplemented by seasonality- and risk-aware adjustments. A key element of the methodology is the separation of predictable seasonal variation from residual volatility, which enables more accurate assessment of the risk associated with individual product items and the use of these risk estimates in working capital allocation decisions. The methodology yields a structured feature vector that includes forecast, statistical and economic characteristics of product items and is suitable both for one-step decision-support systems and as a state-space representation for subsequent stochastic and dynamic inventory control models. The scientific contribution of the work lies in integrating known tools for forecasting, volatility assessment and working capital allocation into a single mathematically consistent methodological model that accounts for the specifics of electronic marketplaces and the particular constraints faced by sellers. The article substantiates the proposed methodology, describes its architecture and presents the design of computational experiments intended for subsequent empirical evaluation of its effectiveness.

Airline schedule planning is a complex multi-stage optimization process, including schedule design, fleet assignment, and aircraft routing. This paper proposes an integrated mathematical model for the joint optimization of fleet type assignment and flight departure times determination within a hub-and-spoke network structure. The model is formulated as a mixed-integer linear program built upon a space-time graph representation of the flight schedule. A key distinguishing feature of the model is the handling of connecting passenger itineraries: revenues from both point-to-point and transfer passengers are included in the objective function, along with fleet operating costs. Problem dimensionality is reduced by exploiting the rotation structure of hub-and-spoke networks: outbound and return legs between a hub and a spoke region are treated as a single rotation unit. The model was validated on real-world data from a major Russian airline, comprising 260 daily flights aggregated into 146 rotations and inter-hub flights, operated by a fleet of 61 aircraft across 5 fleet types. Two independent demand data sources were used for validation: the airline's own booking data and estimates generated via a gravity model. A two-step optimization scheme is proposed: in the first step, departure times are searched over a wider range with a coarser interval; in the second step, the solution is refined over a narrower range with a 5-minute resolution. The optimized schedule achieved a 4–9% increase in forecasted operational profit and a 19% improvement in the number of reachable transfer markets within a three-hour connection window. These results demonstrate that the simultaneous optimization of schedule timing and fleet assignment yields meaningful commercial gains and gives more convenient connection opportunities for passengers.

This paper considers the problem of robust tracking of aerial objects in a video stream, with an emphasis on predicting their trajectories in the presence of measurement noise and short-term detection losses. The aim of the study is to develop a hybrid tracking method combining a classical Kalman filter and a GRU-type recurrent neural network to improve the accuracy and robustness of object motion prediction. The proposed approach is based on the use of a Kalman filter to estimate the current state of an object, specified by the coordinate vector of the YOLO bounding box and the velocities [сх,су, w, h, vx, vy],, and a GRU-type neural network model, which is trained to predict the change in velocity Δvx,Δvy based on a sequence of previous states. Unlike traditional methods, where the motion model is specified a priori, in this paper the motion dynamics are partially extracted from the data. GRU predictions are integrated into the Kalman filter model through an adaptive smoothing mechanism with a coefficient α, which allows for flexible adjustment of the contribution of the neural network prediction to the final state estimate. The main content of the paper includes a description of the hybrid model architecture, a methodology for preparing a training dataset, and an algorithm for integrating neural network predictions into the filtering process. Particular attention is paid to the problem of long-term forecasting—up to 30 frames ahead—which is critical for active tracking and control systems. The scientific novelty of the paper lies in the proposed method for combining GRU and a Kalman filter through adaptive smoothing of the control action, which allows for the consideration of nonlinear and nonstationary characteristics of object motion without sacrificing the interpretability of the classical model. Unlike existing solutions, the proposed method ensures more stable operation under noisy measurements and missed detections. Experiments demonstrated that the proposed approach improves trajectory prediction accuracy and reduces error accumulation during multi-step forecasting compared to the classical Kalman filter and basic neural network models. As a result of the experiments, it was shown that the proposed approach increases the accuracy of trajectory prediction by 16.57% and reduces the accumulation of errors in multi-step forecasting by 21% compared to the classical Kalman filter and basic neural network models. The developed method can be applied to unmanned aerial vehicle detection and tracking, video surveillance systems, and autonomous navigation.

This paper addresses the problem of group control of functionally heterogeneous warehouse robots – loading robots (LR), transport robots (TR), and unloading robots (UR). The aim is to develop a Mivar-based information processing and decision-making system that provides dynamic classification of transport tasks, executor selection via a compatibility criterion, and adaptive fleet energy management while preserving linear inference complexity and full decision explainability. A three-category task classification – individual, cooperative, and collective transportation – is proposed, implemented through Boolean indicators, dynamic priority computation, and dimensional-weight constraint verification. For collective transportation, a spatial clustering mechanism with dual-mode route optimization (exhaustive search and nearest-neighbor combined with 2-opt) is developed. For cooperative transportation of oversized cargo, a combinatorial auction mechanism for TR team formation is introduced, minimizing a composite cost function accounting for payload redundancy, speed heterogeneity, and travel costs. State, operability, and energy sufficiency equations are formulated for each robot type. The mathematical model is transformed into a Mivar knowledge base of 302 IF-THEN rules and verified on  the KESMI platform. The novelty lies in the simultaneous integration of fleet functional heterogeneity, three-category task classification, and adaptive energy management within a unified Mivar knowledge base – aspects previously addressed only in isolation. Experimental validation on a simulation system (1000×1000 m warehouse, 140 robots, 300 tasks) demonstrated a task completion rate of 98.9–99.9%, exceeding baseline Nearest Neighbor and Greedy algorithms by 14–18 percentage points. The method maintains operability under failure of up to 30% of robots and with cooperative task shares reaching 30%.

This article proposes a document structure description model based on Hierarchical Document Structure Analysis (HDSA) methods and the formalism of graph theory. Within the proposed approach, a document is considered as a hierarchical system of interconnected structural elements, including headings of different levels, text blocks, tables, formulas, images, and other types of content. Each element is interpreted as a graph vertex, while nesting and logical dependency relationships between elements are described using directed edges of the graph structure.
The application of the proposed model to a corpus of documents from various subject areas made it possible to identify stable patterns in the organization of textual material characteristic of specific document types. Based on the obtained characteristics, a set of JSON templates was developed to describe typical document structures for different subject areas, including technical, legal, and scientific and educational materials.
Each template contains a formalized description of the document structure, including permissible heading levels, constraints on content types, and parameters of stylistic and compositional formatting. Additionally, a template may include information on the preferred data presentation format, rules for generating tables and formulas, and specialized textual constructions. The use of the JSON format provides a machine-oriented representation of the document structure and enables the integration of templates into software systems for automatic text generation. The developed templates are intended for use in text generation by large language models (LLMs). Within the proposed approach, a template serves as a formalized structural constraint that defines the permissible organization of the resulting document and the sequence in which its sections are generated. This makes it possible to reduce the structural variability of generated text, improve the consistency of interrelated sections, and ensure that the resulting document complies with the requirements of a specific subject area.
The proposed approach is aimed at implementing controlled document generation, in which a large language model generates content within a predefined structural scheme. This provides more stable document generation quality, reduces the likelihood of violations in the logical organization of the text, and improves the reproducibility of generation results. Furthermore, the use of structural templates makes it possible to adapt the generation process to different document types without modifying the architecture of the language model itself.