# Data audit (P1)

| file | SHA-256 (as provided) | data rows | cols | date range | empty cpi | dup dates | month gaps |
|---|---|---|---|---|---|---|---|
| cpi_vintage_2016-07-15.csv | 281fbcdd6a6eeb3e9cb7c1f20476e6bd1f7b925bf140606fe46538fe83f2c83e | 834 | date,cpi | 1947-01 … 2016-06 | 0 | 0 | none |
| cpi_vintage_2026-10-28.csv | 8220d5b4d70ca32fc2c7b1975651c515967db8a87053cf725159c63ed274cfba | 153 | date,cpi | 2014-01 … 2026-09 | 1 (2025-10) | 0 | none |
| cpi_assembled.csv | ff732b201a525d6fb171d4e8b2d470f986863c1547a1928e42bb197c0777d44d | 957 | date,cpi,vintage | 1947-01 … 2026-10 | 1 (dated 2025-11) | 0 | 2016-07 absent |

Units: index 1982–84=100, 3 decimals. Value ranges plausible (21.48 … 336.298).

Notes:
- SHA-256 values in README are of the *raw downloads*; README states files were reformatted, so mismatch with the provided files is expected and not by itself an error (cannot be verified offline).
- Vintage 2026-10-28 has no value for 2025-10 (source reports none; consistent with the BLS non-publication of Oct-2025 CPI). Not imputed: g_2025-10 is missing (and g_2026-10, outside window).
- Overlap 2014-01…2016-06 (30 months): the two vintages differ (seasonal-factor revisions, e.g. 2016-06: 239.927 vs 240.222). Rule: use 2026 vintage.

## Assembly errors in cpi_assembled.csv (file not used for computation; only audited)
1. 2014-01…2016-06 values taken from the 2016 vintage (labelled 2016-07-15) although the 2026 vintage covers them.
2. Month 2016-07 is missing and all 2026-vintage values from 2016-07 onward are shifted one month later (e.g. assembled 2026-10 = 336.298 = vintage 2026-09; assembled 2025-11 empty = vintage 2025-10 empty).
3. Consequently it spuriously extends to 2026-10.
Decision: compute exclusively from the two authoritative vintage files.

No input violates the availability rules (the 2026-10-28 vintage is the latest; nothing outside data/ is used).