Non-invoice documents in cost totals

A folder named for invoices collects whatever looks financially relevant. Every document in it may be legitimate and still not belong in the total.

The situation

A folder assembled over months, holding what everyone involved described as the project’s invoices. Totalling it produced a figure that did not match the ledger, and the difference was not accounted for by duplicates.

The pattern

Three kinds of document had entered the folder without anyone making a mistake.

Payment confirmations restate an invoice that is already in the folder. They carry the same amount and look, to an extraction process, exactly like a second invoice — the double-count from a different direction.

Quotes and pro-formas describe work that was priced but may never have been performed. They carry an amount, a supplier and a date, and nothing in their layout distinguishes them from an invoice except a word most extraction processes never read.

Documents from an entirely different transaction are the expensive case. A closing statement or a loan schedule filed for reference carries a figure orders of magnitude larger than any invoice present. One such document can dominate a total on its own.

The common thread is that filing is not classification. A document is in the folder because someone thought it was relevant, which is a weaker claim than it being an invoice. Classification has to read the document — its structure, its language, the relationship between the figures it carries — and it has to happen before any figure is added to anything.

A practical consequence: sorting by amount descending and examining the largest items finds this class of problem faster than any other check, because the documents that do most damage are the ones that least resemble their neighbours.

Engineering takeaway — Classify document type before any arithmetic. Filing is not classification; classification reads the document. A total is only as trustworthy as the decision about what was allowed into it.