SolarChain-Eval: physics-constrained benchmark for trustworthy economic agents

Researchers propose SolarChain-Eval, a physics-constrained benchmark designed to evaluate the trustworthiness of autonomous economic agents in decentralized energy markets. The framework formulates market governance as a Gymnasium-compatible Markov Decision Process where agents make hourly decisions, assessing policies across dimensions such as market utility, physical safety, and spatial fairness.

• Incorporates an LLM-based Planner/Auditor layer that defines action bounds and reviews high-risk actions.
• Records all interventions through structured logs including trigger signals and audit rationales.
• Experiments with RL and RL+LLM policies reveal a clear utility-safety trade-off.
• Removing physics penalties allows reward-maximizing agents to exploit invalid generation data.

The study indicates that trustworthy agentic AI evaluation requires both physical constraints and transparent intervention traces, with data and code released as open access on GitHub.