Leakage-controlled protocol and confound audit — code and derived data

Everything needed to reproduce the statistical results of the</p><p dir="ltr">article is here.</p><p><br></p><p dir="ltr">WHAT THIS ARCHIVE CONTAINS</p><p><br></p><p dir="ltr">code/pipeline/ the evaluation pipeline: configuration, windowing, features,</p><p dir="ltr"> splits, the seven models, training, the four confound probes,</p><p dir="ltr"> figures</p><p dir="ltr">code/revision/ one standalone script per result, named after the reviewer</p><p dir="ltr"> point it answers, plus verify_numbers.py, which checks every</p><p dir="ltr"> number printed in the article against the result file that</p><p dir="ltr"> produced it</p><p dir="ltr">splits/ the fixed session inventory and the 21 leave-one-session-out</p><p dir="ltr"> folds, as JSON</p><p dir="ltr">derived/ per-window band-energy features with their record, session and</p><p dir="ltr"> label indices; the saved held-out probabilities of all seven</p><p dir="ltr"> deep models; the permutation null</p><p dir="ltr">results/ every JSON produced by the scripts in code/revision</p><p><br></p><p dir="ltr">FIXED SETTINGS</p><p><br></p><p dir="ltr">Seed 42 everywhere unless a script states otherwise.

Window 4,096 samples, hop</p><p dir="ltr">2,048, per-window z-score. Bands 0-200 Hz, 200-1000 Hz, 1-3 kHz, 3-6.132 kHz.</p><p dir="ltr">Deep models: AdamW, learning rate 1e-3, weight decay 1e-4, cosine annealing,</p><p dir="ltr">batch 128, at most 40 epochs, class-weighted cross-entropy, early stopping with</p><p dir="ltr">patience 8 on one inner-validation session drawn from the training sessions of</p><p dir="ltr">the fold.</p><p><br></p><p dir="ltr">WHICH SCRIPT PRODUCES WHICH TABLE OR FIGURE</p><p><br></p><p dir="ltr">Table 3, per-session metadata code/revision/r2_07_session_metadata.py</p><p dir="ltr">Table 5, record and session level code/revision/r2_05_session_level.py</p><p dir="ltr">Table 6, Probe 3 clustered interval code/revision/r2_04_probe3_hierarchical.py</p><p dir="ltr">Table 6, permutation within speed strata code/revision/r2_03_restricted_permutation.py</p><p dir="ltr">Table 7 and Section VI-C, band-restricted repetition</p><p dir="ltr"> code/revision/r2_08_build_lowpass_tensor.py,</p><p dir="ltr"> r2_08_lowband_analysis.py,</p><p dir="ltr"> r2_08_cohen_d_raw.py, r2_08_gpu_rerun.py</p><p dir="ltr">Table 9, multi-seed noise robustness code/revision/r2_09_multiseed_noise.py</p><p dir="ltr">Table 10, public datasets code/revision/download_paderborn.sh,</p><p dir="ltr"> r2_01_paderborn_build.py,</p><p dir="ltr"> r2_01_paderborn_protocol.py,</p><p dir="ltr"> r2_01b_cwru.py</p><p dir="ltr">Figure 6, raw and standardized PSD code/revision/make_figures_revision.py</p><p dir="ltr">Figure 7, novelty-score distributions code/revision/r2_10_novelty_distribution.py</p><p><br></p><p dir="ltr">The full mapping is repeated in the README.md inside the archive.</p><p><br></p><p dir="ltr">REPRODUCING WITHOUT THE RAW RECORDINGS</p><p><br></p><p dir="ltr">derived/band_energy_features.npz holds the per-window feature matrix with its</p><p dir="ltr">record, session and label indices, so every statistical audit in the article can</p><p dir="ltr">be recomputed without downloading the raw recordings.