# P4 checks (2026-10-30)

• Clean re-run (deleted results.json/forecasts.csv/exclusions.csv, re-ran): byte-identical: PASS
• forecasts.csv has 32 rows: PASS
• each origin has 4 rows (2h x 2 models): PASS
• origin < first_release date of targets, target > train_last_date: PASS
• target = train_last + h quarters: PASS
• naive forecast == train_last_value: PASS
• train_last_value matches own vintage last row: PASS
• actual == first_release value: PASS
• RMSE >= MAE > 0 for all: PASS
• rel_improvement consistent: PASS
• exclusion ledger sums: 314-151-3=160=n_analysed: PASS
• first_origin < last_origin: PASS
• n_input_rows match audit (307..314, 318, 13): PASS
• mean abs revision equals mean of per-target revisions: PASS
• analysis.py does not read current_vintage for estimation (only cur.loc in revision line): PASS
• Magnitude: model RMSE (1.49/1.40) below naive (2.01/2.27); plausible given naive inherits full quarter-to-quarter noise (e.g. the 2025Q1 first release of -0.3). PASS