Abstract for: ChatXmile: An LLM-Assisted Closed-Loop Workflow for the System Dynamics Process

System dynamics is inherently iterative, but LLM-assisted modeling raises concerns about uncertainty, opacity, and reproducibility. We introduce ChatXmile, a browser-based method that combines dialogue with LLMs and deterministic, rule-based software to create and refine XMILE models. The goal is to make LLM assistance auditable and repeatable by grounding iterations in technical checks and executable artifacts. ChatXmile implements a closed-loop workflow in which an LLM generates or revises XMILE, and the result is subjected to deterministic conformance and technical checks (XSD/XSLT rule checks, units consistency, structural inspection, and executable trial runs). Diagnostics are fed back to guide iterative refinement. Once models pass these checks, the workflow generates (or updates) structured narratives and HTML/JavaScript simulators that support sensitivity analysis, extreme-condition tests, time-step adequacy checks, and scenario experiments. We demonstrate the workflow using six representative, ideal-typical protocols spanning textbook exercises, simplified benchmark models, a news-based example, and an actively evolving “green economy” model. Across cases, ChatXmile enabled rapid iteration from descriptions or existing models to technically consistent XMILE artifacts and runnable simulators. The protocols show that narratives can support concept-level consistency checking and, in selected cases, enable cross-session reconstruction of equivalent models (identical topology and outputs within numerical tolerance). The demonstrations suggest that a single closed-loop workflow can support three complementary pathways: (i) deterministic technical verification, (ii) simulator-enabled behavioral tests, and (iii) narrative-based consistency checks and reconstruction. ChatXmile does not replace broader model validation in context; rather, it provides auditable scaffolding that reduces friction in model building, testing, and dissemination while keeping modelers in control. Future work will systematize benchmarks and extend support for more complex model types. LLM-assisted XMILE and text generation with deterministic checks and human oversight