# Verification checks (P4)

1. Reproducibility: re-ran out/analysis.py; SHA-256 of results.json (59fa9f46…) and forecasts.csv (c5227b96…) identical before/after. PASS
2. Independent recomputation (integer-index loop, ridge via augmented least squares instead of normal-equation solve): |ΔRMSE|,|ΔMAE| ≤ 1.1e-16 for all three models; n_last=413 matches n_analysed. PASS
3. forecasts.csv: 180 rows, 60 per model; origins 2021-09-01…2026-08-01; every target_date = origin+1 month > origin; max target 2026-09-01. PASS
4. Chronology: in-script assert max training target date ≤ origin at every origin; standardisation stats computed inside the origin loop from training rows only. PASS
5. Exclusion ledger: 417 −2 −2 = 413 = n_analysed. PASS
6. RMSE ≥ MAE > 0 for all models (0.883≥0.699; 0.544≥0.429; 0.561≥0.451). PASS
7. rel_improvement identities recomputed from rmse keys (in script, exact). Magnitudes plausible vs baselines (ridge between mean and naive). PASS
8. No randomness → random_seed null. PASS