2 of 4
Verify the data alone, against the real, messy version.
Because the real records have broken dates and missing years, and a design built on clean examples breaks on the first real one.
Pull ten real examples. Look at the worst three. Design for those.