1# Verification checks (P4)2| # | check | outcome |3|---|---|---|4| 1 | Clean rerun (delete results.json/forecasts.csv/exclusions.csv, rerun) reproduces byte-identical files | PASS (cmp identical) |5| 2 | forecasts.csv has 192 rows, 48 per (model,horizon), no missing actual/forecast | PASS |6| 3 | last_target_date_used <= origin-2 months for every row (asserted in analysis.py) | PASS |7| 4 | last_predictor_date_used <= origin-1 month for every model row (asserted) | PASS |8| 5 | Independent masked re-estimation (GDP>t-2, emp>t-1 set to NaN per origin; out/state/leakage_check.py) reproduces model forecasts | PASS (max abs diff 4.9e-7 = 6-dp rounding in csv) |9| 6 | Exclusion ledger: 345 - (8+2+25) = 310 = n_analysed | PASS |10| 7 | RMSE >= MAE >= 0 for every model/horizon; all RMSE > 0 | PASS (0.814>=0.642; 1.140>=0.872; 0.743>=0.604; 0.902>=0.753) |11| 8 | rel_improvement = 1 - rmse_model/rmse_naive recomputed | PASS |12| 9 | Magnitudes plausible vs baseline: naive h3 error > h1 error (longer gap) | PASS |13| 10 | Chronology: first_origin < last_origin; last_target_date_available (2026-02-01) and last_predictor_date_available (2026-03-01) < last_origin | PASS |14| 11 | data/ unchanged (SHA-256 re-verified after accidental data/out creation was removed) | PASS |15| 12 | Gate pgar_gate.py | see phases.json receipts |16