Closed-Form Calendar Estimation in Deep Load Forecasting: per-window error files, probabilistic runs and superseded records

Companion archive to A. Salam, Closed-Form Calendar Estimation in Deep Load Forecasting: An Information-Controlled Benchmark on Eight National Grids, submitted to Energy and AI (2026). The code, the processed corpus, every result record and the manuscript sources are in the repository https://github.com/salllamye-create/planet-load-forecasting; this archive holds the three folders that are too large for it. results_v2/errors/ (errors_*.zip): one .npz per run with the per-window and per-step losses of every forecast window, the input of every Diebold–Mariano test and model confidence set in the paper. results_v2/prob/ (prob_*.zip): the probabilistic runs, quantile forecasts and their scores for every system, model, horizon and seed. results_v2_discarded*/ (discarded*.zip): records superseded by a logged protocol correction, kept for the audit trail; PLAN_v2.md in the repository says when and why. Every path inside the archives is relative to the root of the repository: clone it, unzip every archive at its root, and the scripts named in README.md regenerate every number, table and figure of the paper. README.md documents the file naming and the fields of the .npz members; MANIFEST.md lists the archives; SHA256SUMS.txt holds their checksums.