# Data audit (P1)
| file | SHA-256 | rows (excl. header) | cols |
|---|---|---|---|
| data/DGS10.csv | 8187f89ab6523b3c964255e1736cfcf0e2ca854ed9e50343e04f4dc2945c5cd7 (matches README and bundle.json) | 5217 | 2 |
| data/README.md | 32b825f0...e561a (matches bundle.json) | provenance text | – |

- Schema: `observation_date` (YYYY-MM-DD, all weekdays 2000-01-03 … 2019-12-31, strictly increasing, 0 duplicate dates; Mon 1044, Tue 1044, Wed/Thu/Fri 1043 rows); `DGS10` percent, 2-decimal (1 bp resolution).
- Missing: 215 rows with empty DGS10 (US holidays, also 2001-09-11/12 closures). No non-numeric values. Range 1.37–6.79%.
- After dropping empties: 5002 observations → 5001 changes (first obs has no change).
- Gaps between consecutive available obs (calendar days): 1d 3899, 2d 58, 3d 889 (weekends), 4d 155 (long weekends). Changes after holidays span >1 trading day; per TASK rules they are kept and attributed to weekday of t (e.g. Tuesday after a Monday holiday) — documented as data issue.
- Extreme changes |Δy|≥25bp: 2001-11-15 +25, 2008-09-15 −27 (Lehman), 2008-11-20 −28, 2009-02-17 −25, 2009-03-18 −51 (Fed QE1 announcement). Genuine market events; kept (no outlier removal); robustness noted in report.
- Info-availability: no input violates the rules; only DGS10.csv used. Weekday means for forecasting use changes dated ≤2009-12-31 only.