Know the universe before you rent a slice of it.
A build begins with the question nobody enjoys: who is excluded? Duplicates come out before the total is reported, never after. The delivered count and the approved count are the same number.
Classification check
Verifying categories against the underlying description rather than trusting a label applied upstream.
Change-only refresh
On a repeat build, reporting what actually moved instead of re-delivering the whole file.
Size banding
Grouping organisations into ranges that the underlying evidence can actually support.
If you have an existing file, we can say what is wrong with it before touching it.