# Data audit
| file | SHA-256 | rows | cols |
|---|---|---|---|
| data/gc5y_daily.csv | 2cdda6ccb2147c8861997166cb0d32cf85ff2744e1e7c0f47eded0f354c2db51 (matches README) | 2498 | 3 |
| data/README.md | 7716f8b58d8f4729504a7713a8a3d651cd2f7298f53f668b377215a5b2af1e34 | – | – |

- Schema: date (str, YYYY-MM-DD, all of them parse), yield_5y (float, percent, range 0.48–3.20), post_cut (int).
- Dates run 2010-01-04 → 2019-12-31. They are strictly increasing, with 0 duplicates and 0 missing values in any column. Gaps between rows are 1–5 calendar days, consistent with weekends and holidays.
- **Critical:** post_cut = 0 for all 2498 rows. n_treated = 0. The README retrieval URL uses end_date=2019-12-31, so the file stops before the 2020-03-04 cut. The flag agrees with the rule date ≥ 2020-03-04 (0 inconsistencies).
- No input breaks the availability rules. The problem is coverage: the treatment window is absent, so the estimand cannot be computed. Decision: report premise_invalid and do not use a proxy window.