# Data audit (P1)

| file | SHA-256 | rows (excl. header) | cols |
|---|---|---|---|
| data/UNRATE.csv | 621ff6fadf50910cb7a21ee20ba610c247cb043b29b4e05c4c1a72fadca2a06c (matches README and bundle.json) | 945 | 2 |
| data/README.md | 62a5a815ca9b63bd6f91fdb176f2e1cecc2faa8ca816c45277b02e8ffd9a70c1 | – | – |

- Schema: `observation_date` (string YYYY-MM-01), `UNRATE` (float, percent of civilian labour force, SA).
- Date range 1948-01-01 to 2026-09-01; 945 consecutive months, no gaps, no duplicate dates.
- Missing values: 1 — 2025-10-01 (empty field). (Consistent with the Oct-2025 CPS not being collected; we do not rely on this explanation.)
- Range 2.5–14.8; mean 5.65. Anomalies: COVID-19 spike, 2020-04 = 14.8 (+10.4 pp from 2020-03), followed by large falls in 2020-05..08. These are genuine observations, kept (no outlier removal; decision made a priori).
- Information availability: the file is the final vintage (revised data); evaluation therefore is pseudo-out-of-sample on revised data, not real-time. Data after each origin are masked in code at every origin. No input violates the rules beyond this documented vintage caveat.

## Consequences of the missing month for the design
- Target 2025-10 is missing for (origin 2025-09, h=1) and (origin 2025-07, h=3) → those pairs kept in ledger with empty actual, excluded from stats → expected 35 evaluated pairs per horizon.
- Origin 2025-10 itself has no observation: naive uses 2025-09 (4.4); model iterates from 2025-09 with h+1 steps.
- Origins 2025-11 and 2025-12: the AR(2)-in-differences state needs the 2025-10 level; it is linearly interpolated from 2025-09 and 2025-11 (both ≤ origin) for the state vector only. Regression rows touching the missing month are dropped from estimation.