RUL Evaluation Audit: Corrective Benchmark Code and Sensitivity Checks

Abstract

Corrective evidence for Training adequacy and evaluation populations in remaining useful life benchmarks: 96 primary fits, two numerical-scale diagnostics and nine additional budget checks. All original prediction prefixes were reproduced exactly and every extended fit stopped by validation patience before the threefold cap. Target-comparison preferences and model-winning sets were unchanged under the budget extension. Separate sensor checks on six fixed predictors provide 72 perturbed test-prediction tables and 600 inference timings; the FD004 reference winner changes under prescribed range-scaled noise or bias. This is an exploratory assessment of one simulated benchmark, not a claim of global convergence or maintenance benefits. This software version supplies frozen corrective training/analysis code, full aggregate tables, figures, literature coding, specifications and reproduction instructions. A local materializer restores the required two-directory layout using the companion generated-artifact dataset. The historical root-level code and tables are retained and explicitly identified as historical. The corrective study supersedes architectural interpretations of a nearly constant baseline, same-task improvement claims based on different native truths, and inference treating deterministic tree repetitions as independent fits. Raw NASA sensor files, literature full texts, manuscript/submission files and credentials are excluded. Historical DOI versions do not contain these corrective experiments.