Verification checks (P4)
- Clean re-run: deleted out/results.json, forecasts.csv, exclusions*.csv; re-ran analysis.py; results.json byte-identical to previous run (cmp). PASS
- forecasts.csv has 32 rows (8 origins × 2 horizons × 2 models). PASS
- Hand check naive h1 RMSE: errors -1.7,1.2,0,-0.5,-2.6,3.3,1.3,-2.9 → sqrt(32.33/8)=2.0103 = rmse_naive_h1. PASS
- Hand check revision: |cur-first| for 2024Q1..2025Q4 = .2,.8,.4,.1,.4,1.0,.4,1.2 → 0.5625 = mean_abs_revision_h1_targets. PASS
- AR(1) OLS cross-checked with statsmodels for last origin: identical coefficients and h=1 forecast 3.2106. PASS
- Chronology: every target first_release_vintage_date ≥ origin (asserted in code); train_last_date < target_date; first_origin < last_origin. PASS
- Exclusion ledger: 314 − 151 − 2 = 161 = n_analysed = n_train of last-origin model rows. PASS
- RMSE ≥ MAE for every model/horizon; all RMSE > 0; rel_improvement = 1 − rmse_model/rmse_naive recomputed. PASS
- Information rule: current_vintage.csv used only in the revision statistic (code inspection). PASS