# Verification checks (P4)

1. Reproducibility: re-ran analysis.py; results.json and forecasts.csv byte-identical (cmp) — PASS.
2. forecasts.csv: 180 rows, 60 per model, horizon all 1, target_date = origin_date + 1 month — PASS.
3. Independent re-implementation (augmented least squares ridge, separate code) with ALL data dated > t replaced by random garbage (×1e3) at each origin: max abs difference vs script forecasts (naive, mean, ridge) 5e-11 (CSV rounding) — PASS; demonstrates no look-ahead leakage incl. scaling statistics.
4. Ridge intercept ȳ equals expanding mean (both average ip_growth 1992-04..t) — PASS (diff < 1e-10).
5. RMSE/MAE recomputed from forecasts.csv match results.json (|diff| < 1e-11); rel. improvements identities hold exactly — PASS.
6. actual column equals ip_growth at target_date — PASS.
7. RMSE > 0 for all models; RMSE ≥ MAE for all models (0.883≥0.699, 0.544≥0.429, 0.561≥0.451) — PASS.
8. Exclusion ledger: 417 − 2 − 2 = 413 = n_analysed — PASS.
9. Chronology: first_origin 2021-09-01 < last_origin 2026-08-01 — PASS.
10. Magnitudes: ridge RMSE (0.561) between mean (0.544) and naive (0.883); plausible for monthly IP growth sd ≈ 1.04 full sample.