# Data audit
- data/firm_panel.csv  SHA-256 a1f1af88b1ab05900155991f2468cce09695af978923e5bc3ef9c28d45683b6f; 2140 data rows x 7 cols (expected 300 firms x 7 years = 2100).
- data/README.md SHA-256 26573e982ab6ca9e7f90d2b89bd2878aa6a670f90da21496bff787b59216b9a2 (codebook).
- Schema: firm_id str (F001–F300, 300 unique), year int 2016–2022, leverage [0, 0.95], cash_ratio [0, 0.74], log_sales (ln thousands USD, ~3–8.9), rating_2025 A/B/C/D (constant within firm), default_next 0/1.
- Missing: only default_next (307 rows, all year 2022 = 300 unique firm-years + 7 duplicate copies). Matches codebook (outcome year 2023 beyond data cut).
- Duplicates: 40 duplicated (firm_id, year) keys; all 40 are exact full-row duplicates -> drop, keep one. After dedup: 2100 rows, 300 firms x 7 years, no conflicting duplicates.
- Information-rule violation: rating_2025 (known 2025; built from realised default history through 2022 and average leverage) -> excluded as predictor (leakage).
- No out-of-range values found for leverage/cash_ratio (both in [0,1]).