Reproduction package: Score, input and model-family dependence of input mixing in open-set bearing diagnosis

This deposit contains everything needed to recompute the tables, figures and reported numbers of the manuscript without retraining a model: per-window scores and saved decision states for the 966 fitted models retained by the study, rerun across every configuration contrast reported in the paper; the corrected window catalogue; the partition index arrays; the analysis and verification code; the derived tables; and the audit reports. Source recordings and raw feature arrays are not redistributed; they remain available from the Paderborn and HUST providers cited in the manuscript. Model checkpoints are omitted because no reported number requires them. The two confirmation probes are separate from the 966 retained models and are included with their 180 and 276 fits respectively, together with the frozen protocol that was hashed before execution and its single pre-execution amendment. Section 2.5 of the manuscript reports a source-domain check that decides which scoring rules may be used as evidence. Every score row is stored, including the rows that fail that check, so the exclusion is auditable rather than asserted. A rule whose source-domain AUROC is at or below chance cannot support a claim about the score family; the manuscript states this explicitly instead of reporting such a rule as if it were informative. Code is released under the MIT licence and the derived tables and reports under CC BY 4.0.