ReflectiChain: Epistemic Grounding in LLM-Driven World Models for Supply Chain Resilience

Global supply chains operating under geopolitical volatility expose both the semantic deficits of traditional Operations Research and Reinforcement Learning (OR/RL) and the physical ungroundedness of pure Large Language Models (LLMs). To address these limitations, we introduce ReflectiChain, an agentic framework that utilizes a Generative Supply Chain World Model (SC-WM) for resilience planning, formulated within a Constrained Partially Observable Markov Decision Process (C-POMDP) framework. By leveraging a topology-aware latent dynamics model, the agent performs multi-step "mental rehearsals" to evaluate candidate strategies. Crucially, the framework incorporates a Double-Loop Learning mechanism that couples test-time reflection-in-action with retrospective reflection-on-action via LoRA updates, enabling autonomous self-evolution of the policy. Validated on the multi-agent Semi-Sim benchmark under adversarial policy shocks, ReflectiChain achieves a Pareto-optimal balance across compliance, resilience, and profitability. The system maintains an 82.3% operability ratio under extreme pressure, demonstrating that latent rehearsal is fundamental for approximating Nash Equilibria in complex, non-stationary networks.