Integrating large language models into Peer-to-Peer energy management for multi-tenant buildings: A guardrail approach to ensuring resilience.

While the integration of Large Language Models (LLMs) into Home Energy Management Systems (HEMS) promises intuitive interfaces, their inherent unreliability poses a significant risk to the stability of Peer-to-Peer (P2P) energy markets. This paper proposes and validates a Planner-Supervisor-Executor architecture that ensures resilient operation through a deterministic guardrail. A benchmark of multiple planner models confirmed that while open-source models (Phi-3 mini, Gemma 7B) and a keyword baseline consistently failed on complex, multi-action commands, Google's Gemini 2.5 Flash achieved the highest workflow accuracy at 89.19%, proving its superior capability in interpreting nuanced user intent. The system's backend couples a day-ahead Mixed-Integer Linear Programming (MILP) scheduler with a real-time Model Predictive Control (MPC) dispatcher to manage stochastic solar generation and optimize community trading via a dynamic P2P pricing mechanism, with fairness validated by the Shapley value and Nash Bargaining Solution. Operational safety is enforced through a three-stage validation process. An initial LLM-based semantic guardrail screens for illogical commands prior to workflow generation, the resulting plan is then verified by the deterministic MARCO supervisor, which effectively intercepts unsafe or malformed workflows. Under these safeguards, the proposed system demonstrates substantial cost savings, increases energy self-sufficiency by up to 64% during the monsoon season, and enables safe "what-if" analyses for residents. Overall, this work provides a validated blueprint for decoupling probabilistic LLMs from deterministic control, proving that robust supervision is a prerequisite for safely deploying AI in critical cyber-physical systems.