# Data audit (P1)

• File: data/UNRATE.csv, SHA-256 621ff6fadf50910cb7a21ee20ba610c247cb043b29b4e05c4c1a72fadca2a06c (matches README).
• 945 data rows + header, 2 columns: observation_date (YYYY-MM-01 string), UNRATE (float, percent SA).
• Range 1948-01-01 .. 2026-09-01, contiguous monthly (no gaps in dates), no duplicate dates, all dated day 1.
• Missing values: 1 (2025-10-01, empty field; consistent with the BLS household survey not being conducted in Oct 2025).
• Values: min 2.5, max 14.8 (2020-04, COVID spike, genuine), mean 5.65.
• Availability: file is final vintage (revised); availability rules are enforced by truncating to dates <= origin in code. No input violates the rules per se; real-time validity is not claimed.
• Missing-value impact: 2025-10 is a target for origin 2025-09 (h1) and 2025-07 (h3) -> those pairs have empty actual and are excluded from stats. Origin 2025-10 itself is missing -> naive uses 2025-09. For the model, interior gaps in the training window (only arises at origins >= 2025-11) are linearly interpolated using only data <= origin; at origin 2025-10 (trailing gap) the model forecasts from the last observed month 2025-09 with steps = target - 2025-09.