FJSP-E2E: A process-verified framework for automatic scheduling-model reconfiguration in smart manufacturing

In smart manufacturing systems, scheduling models must be continuously reconfigured in response to machine failures, process changes, worker availability, tool constraints, AGV logistics, maintenance windows, and quality risks. Large language models (LLMs) provide a natural-language interface for translating shop-floor updates into executable scheduling models, but final objective-value checking is insufficient: an LLM may omit tool, worker, transport, or calendar constraints while still obtaining an accidentally correct makespan on a particular instance. This paper proposes FJSP-E2E, a process-aware framework and benchmark for automatic flexible job-shop scheduling (FJSP) model reconfiguration. FJSP-E2E contains 72 expert-audited diagnostic samples organized into three levels, from explicit instance edits to structural CP-SAT extensions and expert-language compositional reconfiguration. Each sample requires executable Python code and reports both solver results and manufacturing process diagnostic fields. These fields are structural proxies rather than complete formal verification of the CP-SAT constraint graph, but they expose omissions that objective-value checking alone may miss. Experiments on representative LLMs show large differences in deployment readiness: Claude Sonnet 4.5 reaches a 93.1% few-shot pass rate, whereas the weakest model reaches 9.7%, and supervised fine-tuning Qwen3-8B on 4500 synthetic samples improves its pass rate only to 4.2%. The results reveal manufacturing-specific bottlenecks in route choice, AGV transport, worker-resource coupling, and implicit shop-floor language, indicating that reliable LLM-based modeling assistants require process-aware verification rather than answer-only evaluation.