# Verification checks (P4)

1. Clean re-run (deleted results.json, forecasts.csv, exclusions.csv; reran analysis.py): results.json byte-identical — PASS.
2. forecasts.csv has 192 rows, 48 per (model, horizon) — PASS.
3. last_target_date_used <= origin − 2 months for all rows; last_predictor_date_used <= origin − 1 month for all model rows — PASS. Code also asserts max training target month <= t−2 at every origin.
4. actual equals gdp_yoy of target month; target_date = origin + h; naive forecast = gdp_yoy(t−2) — PASS.
5. RMSE >= MAE > 0 for all four series; rel_improvement = 1 − rmse_model/rmse_naive to 1e-12 — PASS.
6. Exclusion ledger: 345 − (10 + 25) = 310 = n_analysed (last origin, h=1 sample) — PASS.
7. Independent re-implementation (dict-based, data physically truncated at t−2 / t−1) for origins 2022-05 h1, 2024-01 h3, 2026-04 h3 reproduces forecasts to 6 decimals — PASS.
8. Magnitudes: naive RMSE 0.81 (h1) and 1.14 (h3) pp, model 0.74 / 0.90 pp; plausible relative to gdp_yoy sd 2.67 over full sample.
9. Chronology: last_target_date_available 2026-02-01 < last_predictor_date_available 2026-03-01 < last_origin 2026-04-01 — PASS.