Trustworthy AI for Quantitative Systems Engineering via Deterministic Delegation to Verified Tools

Abstract

Generative AI offers substantial capability in systems engineering processes for multimodal interaction supporting elicitation, interpretation, and synthesis. However, large language models (LLMs) remain unreliable for exact quantitative computation as they may produce incorrect or hallucinated results when executing mathematical and domain-specific quantitative calculations. Systems engineering is pervasively grounded in quantitative models that demand determinism, correctness, and traceable results, requirements that probabilistic language models cannot reliably satisfy on their own. A deterministic delegation pattern is presented that addresses this limitation by constraining the LLM to input interpretation, assumption documentation, and results presentation while assigning all authoritative computation to verified external tools through formally defined API endpoints. Identical inputs produce identical results, every output carries full provenance, and users can independently inspect or rerun any analysis via a returned source URL. The pattern is demonstrated through application domains for systems engineering cost modeling using the Constructive Systems Engineering Cost Model (COSYSMO) and system reliability analysis using fault tree and reliability block diagram methods. Examples at increasing levels of complexity show how the assistants translate diverse user inputs into complete, reviewable input configurations before invoking the deterministic backend, combining the accessibility of conversational AI with the correctness and auditability that engineering decisions require.