# Data audit (P1)
| file | SHA-256 | rows | cols |
|---|---|---|---|
| data/canada_gdp_employment.csv | f4c8e4976fb5f5f0c326e40bcb4f85b8b22963e5be89e2729e1a430c198f2bed (matches README) | 345 data rows + header | 5 |
| data/README.md | 2000e2e64c856f80d0781666df479d4ebfc63cb09ce41d9db99c04ca2b92fa2f | — | — |

Schema: month (YYYY-MM-01, 1998-01..2026-09, monthly, contiguous, no duplicates); gdp_level (chained 2017 $ millions, SAAR); gdp_yoy (%); emp_level (thousands); emp_yoy (%).
Missing: gdp_level/gdp_yoy missing only for 2026-08, 2026-09 (not yet published) — expected. emp complete.
Recomputing yoy from levels reproduces given columns to <5e-5 (rounding to 4 decimals).
Anomalies:
- README says gdp_yoy is "Empty for the first 12 months", but the file has gdp_yoy (and emp_yoy) for 1998 (presumably computed from 1997 levels outside the file). Harmless; values used as given.
- COVID-19 shock: gdp_yoy reaches -16.2 (2020-04) and +18.5 (2021-04, base effect); emp_yoy similar. These extreme points would dominate OLS. Decision: exclude from model *training* any row whose target or predictor month falls in 2020-03..2021-12 (decision made using only pre-2022 information, i.e. available at the first origin 2022-05). Naive forecast unaffected.
Information rules: the file is a single current-vintage table; availability is enforced in code by truncating gdp to months <= t-2 and emp to <= t-1 at each origin. No input violates rules per se; the vintage issue (revisions) cannot be addressed and is reported as a limitation.