For high-stakes AI applications, classical expert systems (XPS) offered transparent, rule-based reliability, whereas modern large language models (LLMs) provide broad knowledge but often lack explainability and consistency. This paper presents XPS 2, a neuro-symbolic architecture that evolves the classic diagnostic expert system by integrating a powerful LLM under strict symbolic control. In XPS 2, the LLM only proposes hypotheses, which are rigorously tested and verified by a rule-based controller using formal domain rules and trusted external tools, ensuring no unverified output influences decisions. We outline key design principles and a multi-layer system architecture (control, intelligence, knowledge, and tooling) that collectively enforce thorough validation, explicit state tracking, and complete audit trails of the reasoning process. We also compare XPS 2 against conventional tool-augmented LLMs and fully autonomous LLM agents, highlighting how XPS 2 provides clearer decision authority separation, systematic result verification, and safer failure modes. Initially demonstrated for technical fault diagnosis tasks (e.g., complex equipment troubleshooting), the XPS 2 approach is extensible to other critical domains such as automotive, healthcare, or aerospace where dependable and explainable AI support is essential.
Roman R. Laczkovich (Thu,) studied this question.