Causally Grounded LLM Attribution Agents for High-Dynamic Logistics Systems: Design and Experimental Validation

High-dynamic logistics systems frequently generate anomalies due to interacting operational mechanisms like demand surges, driver shortages, and exogenous shocks. While large language models (LLMs) can transform heterogeneous telemetry into natural-language explanations for operator diagnosis, unconstrained language reasoning remains unreliable for root-cause attribution in systems with structured dependencies. To address this, we propose a causally grounded attribution agent architecture integrating a streaming state-preparation layer, a structural causal graph (SCG) to constrain admissible cause-effect paths, a quantitative attribution core, and an LLM reasoning layer. This framework converts grounded evidence into reliable explanations and intervention suggestions. We validate the core components on a controlled synthetic benchmark. The SCG-aligned model achieves a superior macro F1 score of 0.753 on the in-distribution test set and demonstrates robust performance under distribution shifts, outperforming random forest and ungrounded heuristic baselines. Furthermore, a graph misspecification study confirms that the SCG provides critical structural information beyond mere regularization, as removing a single causal edge significantly reduces accuracy. Finally, an LLM evaluation across multiple grounding configurations reveals that full causal grounding improves attribution accuracy by 20 to 35 percentage points, with smaller models benefiting disproportionately. Ultimately, this study contributes a robust, causally grounded agent architecture and a replicable cross-tier evaluation framework for LLM-based causal reasoning, laying the groundwork for future validation on production telemetry and downstream operational impact assessments.