# Data audit (P1)

Schema: date (YYYY-MM-DD string, parsed fine), 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.
• Rows per year: 249–250 for each of 2010–2019 (business-daily). Gaps between consecutive rows: 1–5 calendar days (weekends/holidays), no long gaps.
• yield_5y: min 0.48, max 3.20, mean 1.574 %; plausible.
• post_cut: all 2498 rows are 0; n_treated = 0. The README's retrieval URL used end_date=2019-12-31, so the file stops ~2 months before the 2020-03-04 announcement. The flag is consistent with its definition (no dates ≥ 2020-03-04 exist), so it is not mis-coded — the post period is simply absent.

Consequence. The estimand (post-cut mean change minus mean change in the 120 rows before the first post_cut=1 row) is not computable: there is no treated window and the pre-window is defined relative to a first flagged row that does not exist. Per TASK.md, conclusion = premise_invalid, feasibility = infeasible, estimand keys null. No proxy (e.g., redefining the treatment date) will be substituted. Descriptive baseline (whole-file mean one-day change), table and figure are still produced.

No information-availability violation in the inputs (only the single permitted file is used).