TRACE Framework for Trustworthy Agentic AI in Critical Domains
TRACE serves as an engineering framework aimed at ensuring trustworthy agentic AI across critical operational domains, such as clinical decision support, industrial multi-domain operations, and judicial AI assistance. It features a four-layer reference architecture that differentiates between classical machine learning and LLM validation (L2a/L2b), a policy for stateful orchestration and escalation (L3), and incorporates limited human oversight (L4). Additionally, it presents a trust-metric suite grounded in metrology, adhering to GUM, VIM, and ISO 17025 standards, and introduces the Computational Parsimony Ratio (CPR) as a key design principle. The L2a/L2b distinction emphasizes intentional design choices for LLM usage. The framework is detailed in a paper available on arXiv within the computer science and computation and language categories.
Key facts
- TRACE is a cross-domain engineering framework for trustworthy agentic AI.
- It targets operationally critical domains: clinical, industrial, judicial.
- Architecture has four layers with a classical-ML vs LLM-validator split (L2a/L2b).
- Includes stateful orchestration-and-escalation policy (L3) and bounded human supervision (L4).
- Trust-metric suite is metrologically grounded to GUM, VIM, ISO 17025.
- Introduces Computational Parsimony Ratio (CPR) as a design principle.
- Three instantiations: clinical decision support, industrial multi-domain, judicial AI assistant.
- Paper available on arXiv, submission history and browse context provided.
Entities
Institutions
- arXiv