# Data audit

• data/firm_panel.csv SHA-256 a1f1af88b1ab05900155991f2468cce09695af978923e5bc3ef9c28d45683b6f (matches bundle.json); 2140 data rows, 7 columns.
• data/README.md SHA-256 26573e982ab6ca9e7f90d2b89bd2878aa6a670f90da21496bff787b59216b9a2.
• Schema: firm_id str (F###, 300 distinct), year int 2016–2022, leverage/cash_ratio (ratios, 0–1), log_sales (log thousands USD), rating_2025 (A/B/C/D, constant per firm), default_next 0/1.
• Missing: only default_next, 307 rows, all year 2022 (outcome year 2023 not observed at data cut). 0 missing elsewhere.
• Duplicates: 40 exact full-row duplicate rows (40 firm-year keys appear twice with identical values). After removal: 2100 rows = 300 firms x 7 years, balanced.
• Ranges: leverage 0.02–0.95; 57 firm-years at exactly 0.02 (apparent floor/censoring in the simulated data) — kept, documented. cash_ratio 0.0046–0.7406; log_sales 1.29–8.88. No out-of-range values.
• Availability violation: rating_2025 known in 2025 and built from realised default history 2016–2022 (includes the outcome). Default rate in years<2022 is 0 for ratings A/B/C and 0.227 for D => leakage. Decision: exclude from analytic dataset, list in variables_excluded.
• Identifiers firm_id/year are not predictors; default_next is the outcome.