Alarm-Policy Sensitivity in Bearing Endpoint Warning: A Conveyor-Motivated Dataset Audit

This preprint presents a reproducible audit of alarm-policy sensitivity in bearing endpoint warning. The study examines how different definitions of warning success—particularly newly issued alarm episodes versus persistent active alarm states—can lead to substantially different conclusions even when the underlying model scores remain unchanged.Experiments are conducted on complete XJTU-SY and PRONOSTIA bearing histories using five fixed comparators, whole-bearing separation, training-only transformations, horizon and threshold sensitivity analysis, and causal alarm persistence rules. Results show that apparent event coverage can increase markedly under state-overlap evaluation while notification burden and early alarm frequency remain substantial.The work emphasizes joint reporting of timely alarm onset, active-state overlap, notification workload, state occupancy, and endpoint construction. It is intended as an auditable methodological case study rather than a claim of a deployable conveyor predictive-maintenance system.The accompanying reproducibility package includes configurations, source manifests, extracted features, scores, predictions, fold parameters, and computational verification scripts. Raw third-party datasets are not redistributed and remain subject to their original access and reuse terms.