Reproducibility package for: Sensor-Modality Robustness Under Speed Shift — A Pre-Registered Evaluation of the Rotating Electromechanical System Benchmark

Complete artifacts for a pre-registered evaluation of the Rotating Electromechanical System dataset (DOI 10.34810/data2500, CC BY 4.0), comparing five classical feature families with gradient-boosted trees against two published deep architectures across six evaluation protocols, with statistical inference at the level of held-out environments rather than segments or seeds. Contents: Pre-registration (PREREG.md) with dated amendments A1, A2, A(e), A(f), A(g) and A(h), each stating its interpretation rule before the corresponding analysis was executed Results manifest: one JSON record per training–evaluation unit (1,240 units), carrying the full configuration, winning hyperparameters, all metrics, per-class F1 and confusion matrix Frozen segment definitions and fold assignments for all six protocols (P1, P1R, PSR, P2S, P2T, P3) Statistical summaries Published figures and the analysis notebooks that generate them Exported requirements file pinning package versions SHA-256 manifest covering every deposited file Running the analysis notebooks against the deposited manifest regenerates every figure, table and in-text value in the article without retraining. Reproducing the manifest itself additionally requires the raw record from the dataset DOI above; the raw record is not redistributed here. Chronology. The experimental blocks were executed between 17 and 31 July 2026, with the offset-invariant thermal arm A(h) on 12 August 2026. This deposit was created on 1 August 2026 and updated thereafter, so its timestamps postdate the work and do not by themselves evidence that the pre-registration preceded the experiments. What the deposited pre-registration does show is the content of each amendment and the interpretation rule fixed for it, including rules whose outcomes weakened headline claims: A(g) declared two branches in advance and branch (ii) fired, requiring the thermal attribution to be reported as an upper bound; A(f) changed the primary inferential unit after results existed, downgrading several comparisons that had been significant. Ambient control, A(g). Deposited as two records in verification/: an initial run using a segment-level permutation null (empirical p = 0.0952), superseded by the acquisition-level null reported in the article (p = 0.0249). The registered criterion was the 95th percentile of the null rather than a p-value threshold, and the observed value exceeds it under both nulls, so the outcome is unchanged. Both records and both nulls are retained, and the deviation and its reason are stated in amendment A(g). Offset-invariant thermal arm, A(h). Deposited with its amendment text and its 45 unit records. Funded by the Deanship of Scientific Research, King Abdulaziz University, under grant IPP: 518-611-2026.