This article introduces a comprehensive framework for building trustworthy and interpretable Large Language Model (LLM)-driven control planes for autonomous distributed systems. As LLMs are increasingly integrated into critical infrastructure management, from database administration to resource allocation, they introduce unique challenges related to trustworthiness, interpretability, and robustness. To identify two fundamental issues: the Recursive AI Problem, where AI systems manage other AI components, creating complex feedback loops, and the Semantic Gap Issue, where LLMs operate at high levels of abstraction while distributed systems require precise operational semantics. Our proposed framework addresses these challenges through a multi-layered approach combining formal verification techniques, explainability mechanisms, and human-in-the-loop integration. We develop novel methods for system-specific LLM fine-tuning, symbolic-neural integration, and semantic alignment verification. Experimental evaluation across database management, Kubernetes resource allocation, and data pipeline orchestration scenarios demonstrates significant improvements in trustworthiness and interpretability while maintaining performance benefits. This work bridges the critical gap between advanced LLM capabilities and the requirements for safe, reliable system management in mission-critical environments.
Shalini Katyayani Koney (2025) studied this question.