What a citation must carry to survive review
Every data vendor claims “sources.” One test cuts through all of it: hand your angriest reviewer a number and leave the room. If they can pull the document without calling you, you win.
There is a moment in every serious review when someone stops reading your conclusions and picks one number to attack. Credit committee, securitization diligence, litigation support, it is always the same moment. The next thirty seconds decide how your next hundred numbers get treated. If the number opens to a document the reviewer can pull themselves, the audit moves on. If it opens to “our data partners,” everything you submitted just became homework.
The four fields that make a citation real
A citation that survives has to carry enough for independent retrieval. In practice that is four things, and the absence of any one of them is disqualifying:
- The source system, named precisely. Not “county records” but
nyc_acris,fl_or,ny_dos_corps. The office and dataset, named the way the office names it, so the reviewer knows exactly which index to open. - The record id. The document number the source itself uses, so the same instrument can be pulled from the same index with no interpretation step.
- The recorded date. When the filing entered the public record. Not when we ingested it. Conflating the two is how staleness hides.
- The epistemic status. Read off a filing, derived by arithmetic, or inferred with a confidence score. Three different kinds of claim, and a reviewer is entitled to know which one they are holding.
Per field, or it does not count
The industry’s favorite compromise is the report-level citation: a sources appendix stapled to a PDF. It fails the thirty-second test structurally, because it cannot say which record produced which number. Provenance has to be carried per field, through every join and every export, or it degrades into decoration. This is an architecture decision, not a formatting one. Provenance bolted on at render time is already broken, because the pipeline beneath it stopped tracking the linkage.
It is also why 21.4M owner records in the corpus carry a source and a field-level map. Not because anyone reads millions of citations, but because any one of them might be the one that gets attacked.
The refusal that builds trust
The last property of a survivable citation is the strange one: sometimes it must refuse to exist. A current loan balance cannot be cited from public records, because no public record states it. When a system declines to print a number it cannot support, every number it does print gets more credible. The vendors who fill that gap with an unlabeled estimate are borrowing credibility from documents their number is not actually in.
The test is not “does it have sources.” The test is: hand your fiercest reviewer one number, and leave the room.