# Data audit (P1)
| file | SHA-256 | rows x cols |
|---|---|---|
| data/canada_gdp_employment.csv | f4c8e4976fb5f5f0c326e40bcb4f85b8b22963e5be89e2729e1a430c198f2bed (matches README & bundle.json) | 345 x 5 |
| data/README.md | 2000e2e64c856f80d0781666df479d4ebfc63cb09ce41d9db99c04ca2b92fa2f | documentation |

Schema: month (date, YYYY-MM-01, 1998-01..2026-09, monthly, no gaps, no duplicates); gdp_level (float, chained 2017 $ millions SAAR); gdp_yoy (float, %); emp_level (float, thousands); emp_yoy (float, %).
Missing: gdp_level/gdp_yoy empty for 2026-08, 2026-09 only (unpublished). emp complete.
Checks: gdp_yoy and emp_yoy recomputed from levels (100*(x/x_{-12}-1), 4 dp) match exactly from 1999-01 on (max abs diff 0.0).
Anomalies: (1) README says yoy columns empty for first 12 months, but file has 1998 values (presumably computed from pre-1998 levels) — used as given. (2) COVID swings 2020-03..2021-12 (gdp_yoy −16.2 to +18.5) — excluded from model training (pre-specified, not tuned on test errors).
Information availability: the file contains data beyond what is available at each origin (current vintage, all months). Handling: the script slices per origin (gdp ≤ t−2, emp ≤ t−1, training targets ≤ t−2); verified by a perturbation test. Vintage issue (revised data) cannot be fixed: reported as limitation, no real-time validity claimed.