An empirically grounded modelling architecture for rapid ex ante policy evaluation

Abstract

Circular economy (CE) transitions emerge from the interplay of heterogeneous consumer and firm behaviours, infrastructure dynamics, and policy interventions. However, the tools widely used to assess CE policy, e.g. LCA, MFA, and EEIO models, are static and cannot represent how markets form and evolve over time. Agent-based models (ABMs) can capture these dynamics, but specifying realistic agent behaviours remains a persistent bottleneck. We propose a hybrid architecture, GABM/SD+EEIO, that addresses this bottleneck by using large language models (LLMs) as structured elicitation tools rather than as runtime decision-makers. The LLM reads a natural-language policy description, identifies how it affects different consumer segments, and selects best-fit parameter values from a curated library of empirical research, assessing their transferability to the modelled context. The ABM then systematically explores the parameter space, producing results with quantified uncertainties. The model is explicit, auditable, and reproducible without LLMs. We validate this approach by experiments showing that LLMs reliably identify the ordinal structure of consumer heterogeneity but cannot replace empirical data for cardinal calibration. A demonstration model for electronics recycling policy illustrates the framework. The architecture is modular and extensible to other policy domains where behavioural dynamics interact with physical and economic systems. A more general successor framework is under development.