# Data audit (P1)
| file | sha256 | rows (excl. header) | cols |
|---|---|---|---|
| data/firm_panel.csv | a1f1af88b1ab05900155991f2468cce09695af978923e5bc3ef9c28d45683b6f | 2140 | 7 |
| data/README.md | 26573e982ab6ca9e7f90d2b89bd2878aa6a670f90da21496bff787b59216b9a2 | codebook | — |
SHA-256 matches bundle.json.

Schema: firm_id (str, F001–F300, 300 distinct), year (int 2016–2022), leverage (ratio, 0.02–0.95), cash_ratio (ratio, 0.0046–0.7406), log_sales (log thousands USD, 1.29–8.88), rating_2025 (A/B/C/D, constant within firm), default_next (0/1, 307 missing).

Findings:
- 40 exact duplicate rows (all 7 fields identical); no conflicting firm-year duplicates. After dedup: 2100 = 300 firms x 7 years, balanced.
- default_next missing for all 2022 rows (307 raw, 300 unique) and nowhere else — right-censoring at the end-2022 cut. Excluded, not imputed.
- rating_2025: known in 2025 and built from realised defaults through 2022 + average leverage -> violates info rule -> column excluded.
- leverage has a mass point at 0.02 (57 raw rows) — likely floor in generation; kept.
- No missing predictors; no out-of-range ratios (all in [0,1]).
- Values in default_next: 1609 zeros, 224 ones (raw, incl. duplicates).