Nothing is assembled before the target is described in writing. Anything inherited from a source is tested rather than assumed. Columns without a written meaning are left out rather than shipped undocumented.
Match confidence
Separating confident pairs from probable ones, and reporting the latter rather than folding them in silently.
Residual questions
The parts of a brief the data cannot answer, returned as questions rather than guesses.
Outlier review
Looking at the extremes of each field, because that is where bad data announces itself.
Send the field list and we will mark which parts are reachable.