# Data audit (P1)
## data/DGS10.csv
- SHA-256 8187f89ab6523b3c964255e1736cfcf0e2ca854ed9e50343e04f4dc2945c5cd7 (matches data/README.md and bundle.json).
- 5217 data rows + header; 2 columns: observation_date (YYYY-MM-DD), DGS10 (percent, 2 decimals → 1 bp resolution).
- Dates 2000-01-03 … 2019-12-31, strictly increasing, no duplicate dates, weekdays only (Mon 1044, Tue 1044, Wed 1043, Thu 1043, Fri 1043).
- Empty DGS10: 215 rows (holidays). No non-empty unparsable values. Non-missing levels: 5002, range 1.37–6.79 %.
- Changes: 5001, mean −0.093 bp, sd 5.72 bp, min −51 bp (2009-03-18, FOMC QE1 announcement), max +25 bp. Extremes are genuine market events; kept.
- Gaps between consecutive available obs (calendar days): 1 (3899), 3 (889, weekends), 4 (155, long weekends/holidays), 2 (58, mid-week holidays). Hence some changes span >1 trading day (post-holiday); by rule they are attributed to the weekday of t. Notably, after Monday holidays the Tuesday change spans Fri→Tue.
- data/README.md: provenance text only, not analysed.
## Information-availability
No input violates the rules. Only DGS10.csv used. Training weekday means use only changes dated ≤2009-12-31; the first 2010 change (dated 2010-01-04) uses the 2009-12-31 level, which is permitted because the change is dated in the evaluation period.
## Exclusions
Only the rule-mandated drop of 215 empty rows plus the first available observation having no change (5217 → 5002 levels → 5001 changes).