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

- Schema: date (str YYYY-MM-DD), yield_5y (float, percent), post_cut (int).
- Date range 2010-01-04 → 2019-12-31, strictly increasing, 0 duplicate dates, 0 missing values in any column.
- yield_5y: min 0.48, max 3.20, mean 1.574 %, plausible.
- Calendar gaps between rows: 1d ×1962, 2d ×10, 3d ×440 (weekends), 4d ×79, 5d ×6 (holiday stretches) — consistent with business-daily data; no long outages.
- Rows per year: 249–250 for each of 2010–2019.
- **post_cut: 2498 zeros, 0 ones.** The README states the retrieval was end_date=2019-12-31, so the file contains no dates on or after 2020-03-04. The treated window is empty and the 120-day pre window (defined relative to the first flagged day) is undefined.
- Availability rules: no violation in the inputs (no future information); the problem is missing coverage, not leakage.
- Decision: estimand not computable → `premise_invalid`, feasibility `infeasible`; effect, CI, model RMSE and n_analysed set to null. Do NOT substitute e.g. the last 120 days of 2019 or a different event date (that would be a proxy/redefinition of post_cut). The whole-file baseline mean change is still reported as required.