Abstract
Large language models can interpret natural-language goals and formulate quantitative relationships; however, autonomous generation often leads to incomplete variable sets, hidden assumptions, and non-reproducible numerical results, especially under domain-specific constraints. This paper proposes a hybrid architecture that combines LLM-based construction of a domain-specific mathematical model with a deterministic computation layer. The system derives structured variables and executable expressions from a task description, validates inputs against domain rules, retrieves missing quantities from external knowledge sources when needed, and computes formulas in a controlled environment separated from probabilistic decoding. An end-to-end workflow is presented and experimentally evaluated on a diverse set of user tasks against a “LLM-only” baseline. We report task accuracy, input completeness, end-to-end pipeline latency, and the rate of partial outputs. The proposed approach improves accuracy from 72% to 97% and input completeness from 60% to 93%, while reducing the share of tasks with missing inputs from 40% to 4% and partial outputs from 28% to 3%; mean execution time is slightly higher than the baseline (4.2 s vs 2.3 s) due to validation and context retrieval overhead. The results indicate that separating semantic reasoning from deterministic execution improves the reliability of domain-specific mathematical modeling. Key words: computation, deterministic execution, domain-specific modeling, hybrid system, large language models, mathematical model, validation.