Reproducibility Archive (v3): Bridging Energy Optimization and Distributed Information Systems — Five-Layer Taxonomy and Research Agenda

Version 3 of the reproducibility archive for the dissertation on AI-driven data-centre load shifting. This version restores the executable framework code and the validated empirical result tables that version 2 had withdrawn. Contents: (a) the forecaster of record (forecaster_upgraded.py; causal LSTM-Transformer, hidden size 96, 309,224 parameters) and the MILP optimisation layer (PuLP/CBC: milp_static.py, phase5_with_dis.py, dis_filter.py) with experiment drivers; (b) public market/grid datasets (CAISO OASIS LMP; EIA hourly demand for the CISO balancing authority; EIA fuel-mix carbon intensity) and acquisition scripts; (c) validated result tables and prediction files; (d) a README with one-command reproduction and a checksummed MANIFEST. Verified results (reproducible from the included code and data): forecaster point R2 = 0.7454 and 50-seed R2 = 0.7167 +/- 0.0074 (persistence 0.6896, seasonal 0.5147); MILP + BESS + DS-filter 60-day hourly rolling-horizon run: 5.14% cost saving, 3.96% CO2 saving, 0 of 6,799 jobs missed, migration cost 3,183.80 USD (post-hoc accounting, not part of the solved objective). Workloads are synthetic, sampled from assumed state-type and jurisdiction priors defined in dis_filter.py (fixed seed); they are NOT derived from any production trace. Earlier unintegrated exploration scripts are retained, labelled, under deprecated/. This version supersedes v2; cite the concept DOI 10.5281/zenodo.19296694 for the latest version.