# Data audit

• data/firm_panel.csv SHA-256 a1f1af88b1ab05900155991f2468cce09695af978923e5bc3ef9c28d45683b6f (matches bundle.json); data/README.md 26573e98...a6a670f9 (codebook).
• 2140 data rows x 7 columns (firm_id str, year int, leverage, cash_ratio, log_sales float, rating_2025 A/B/C/D, default_next 0/1/empty).
• 300 firms (F001–F300 format), years 2016–2022. Intended 2100 rows.
• Duplicates: 40 exact duplicate rows (= 40 duplicated firm-year keys, identical on all columns). After removal every firm has exactly 7 years.
• Missing: only default_next (307 raw / 300 unique rows), all and only origin year 2022 (outcome 2023 beyond data cut).
• Ranges: leverage 0.02–0.95 (54 analysed rows at 0.02, apparent floor), cash_ratio 0.0046–0.74, log_sales 1.29–8.88. No negatives.
• rating_2025: constant within firm; only D-rated firms ever default in raw data → outcome leakage. Violates availability rule (known 2025) → excluded.