# Data audit (P1)

All SHA-256 hashes computed with `shasum -a 256` match `data/README.md` exactly (23 files). Row counts (data rows after header) match README.

| file | rows | cols | range | missing | dup dates |
|---|---|---|---|---|---|
| PAYEMS vintages 2025-10-31 … 2026-07-31 | 1040,1041,1043,1044,1045,1046,1047,1048,1049,1050 | 2 | 1939-01 → 2025-08 … 2026-06 | 0 | 0 |
| RSAFS vintages 2025-10-31 … 2026-07-31 | 404,405,406,407,408,409,411,412,413,414 | 2 | 1992-01 → last(V) | 0 | 0 |
| RSAFS_current.csv | 417 | 2 | 1992-01 → 2026-09 | 0 | 0 |
| UMCSENT.csv | 887 | 2 | 1952-11 → 2026-09 | 210 empty (pre-1978 non-survey months; monthly without gaps from 1978) | 0 |
| first_release.csv | 28 | 5 | 2024-06 → 2026-09 | 0 | 0 |

Units: RSAFS millions USD SA; PAYEMS thousands SA; UMCSENT index 1966Q1=100; yoy in percent (log ×100).

## last(V) per origin (RSAFS vintage)
2025-10-31→2025-08 (2-month lag, shutdown delay); 2025-11-30→2025-09; 2025-12-31→2025-10; 2026-01-31→2025-11; 2026-02-28→2025-12; 2026-03-31→2026-01 (2-month lag); 2026-04-30→2026-03 (two months added at once); 2026-05-31→2026-04; 2026-06-30→2026-05; 2026-07-31→2026-06.
Consequently h=1 targets: 2025-09,10,11,12, 2026-01,02,04,05,06,07 (2026-03 is never an h=1 target); h=3 targets: 2025-11,12, 2026-01,02,03,04,06,07,08,09. All 20 targets exist in first_release.csv.

## Anomalies
- PAYEMS vintage 2025-12-31 adds two months at once (2025-10 and 2025-11) — delayed BLS publication; PAYEMS vintage 2026-02-28 shows a level shift of ~ −1,000k vs prior vintage (annual benchmark revision). PAYEMS is not used by my model, so no handling needed beyond documentation.
- first_release.csv: yoy recomputes from its levels to < 5e-7; levels match the shipped vintages for the 10 target months whose first-release vintage is shipped. 2026-02 and 2026-03 share first-release vintage 2026-04-30.
- RSAFS revisions between consecutive shipped vintages are small (≤0.42% in level, 4–5 months changed). Current vintage vs vintage 2026-07-31: 138 months changed back to 2015-01, max 4.6% level revision (annual benchmark revision lowering 2025–2026 levels) — relevant for estimand (d) only.
- RSAFS_current.csv contains information after all origins → used only for (d). first_release.csv is used only as actuals.

## Availability violations
None in the forecasting inputs as used: each forecast reads only the RSAFS vintage dated V (my model uses no other predictor). UMCSENT not used by any forecast.