Rethinking replication in machine learning comparisons of CGM glucose forecasting: a two-axis variance decomposition of seed and cohort effects - reproducibility package

Reproducibility package for the manuscript submitted to BMC Medical Informatics and Decision Making. Contents: analysis scripts (Python 3.11), the complete 200-run seed ledger (4 architectures x 20 seeds x 2 cohorts = 160 audit runs; 4 x 5 x 2 = 40 per-subject probe runs), the per-subject NumPy archives underlying the two-axis variance decomposition, and SHA-256 checksums for every file. Source cohorts are public and are NOT redistributed here: - MetaboNet (T1D, 1,092 subjects): https://metabo-net.org - ShanghaiT2DM (T2D, 100 subjects): figshare, as cited in the manuscript Only derived per-subject aggregates are included. Code is MIT licensed; derived results and figures are CC BY 4.0.