TRACE: A Metrologically-Grounded Engineering Framework for Trustworthy Agentic AI Systems in Operationally Critical Domains

Trustworthy AI requires deliberate architectural choices about when to use LLMs: TRACE's key innovation is separating classical ML (L2a) from LLM components (L2b), ensuring that the use of large language models is a conscious design decision rather than a default approach. This separation allows teams to choose the right tool for each task rather than applying LLMs universally.

Trust in AI systems should be measured using established scientific standards: The framework grounds its trust metrics in metrological standards (GUM/VIM/ISO 17025)—the same rigorous measurement principles used in physics and chemistry—rather than ad-hoc metrics. This brings scientific rigor to evaluating whether AI systems are actually trustworthy.

Simpler AI models should be preferred when they perform adequately: TRACE introduces the Computational Parsimony Ratio (CPR) as a formal design principle, quantifying when simpler models are preferable to complex ones. This directly addresses the tendency to over-engineer AI solutions and helps teams justify using less computationally expensive systems.

Human oversight must be structured and bounded, not ad-hoc: The framework includes explicit layers for stateful orchestration (L3) and bounded human supervision (L4), meaning human involvement is designed into the system architecture with clear escalation policies rather than treated as an afterthought.

The same trustworthy-AI framework can work across very different domains: TRACE demonstrates transferability across clinical, industrial, and judicial applications, suggesting that core principles for responsible AI—measurement rigor, architectural clarity, and parsimony—are domain-agnostic.