Nobody Could Say Which Document Was The Real One

A deal team had more documents than it could hold in its head, and no way to tell which copy of anything was the real one. A long management presentation as a scanned file. A seller hosted document room where some folders simply were not visible to the buy side. Spreadsheets with tabs hidden from the viewer and formulas resolving to errors. Attachments scattered across a mailbox. Everyone was reading the same deal and no two people were certain they were reading the same version of it.

The cost of that was not hypothetical and it surfaced in the first week. One of the central commercial terms was being carried verbally, from one conversation, and the whole team had internalised it. When the actual document was read, the term was structured differently. Everyone had been reasoning from a shared belief that was wrong, and nobody could have caught it, because there was nothing to catch it against.

So the first deliverable was not analysis. It was an index. Every artifact in the deal got a number, a type, a pointer to where its machine readable extraction lives, and a verification status: verified, in hand but not yet read, or superseded. That last category is the one that does the work. A source that has been replaced stays in the index, marked, with the thing that replaced it, so nobody quietly re-derives a conclusion from a document that has been overtaken.

The index is also honest about what it has not read. A meaningful share of entries sit at in hand but not yet read, and they are visible as such rather than folded into a total that implies full coverage. An index that claims completeness is worse than no index, because it stops people looking.

The extraction underneath is deliberately mechanical. The presentation was transcribed page by page and then spot checked against the original, including both of the dense financial tables, and the spot check matched exactly. Spreadsheets were dumped programmatically to values only files rather than read through a viewer, which is how the hidden tabs came back: twenty of them, in one workbook, carrying content the on screen version simply did not offer. Image only pages were reviewed visually rather than skipped.

Then the boring, load bearing part. The document room was mirrored locally, and each fresh export was compared against the mirror file by file. The most recent comparison came back with sixty five of sixty five files identical, which is a useful answer and an unglamorous one: nothing new had been posted. Knowing that with certainty is worth more than assuming it.

The last piece is a small script that takes a finished brief and greps every quoted line back to the extracted source text, run with a control so a pass means something. Sixty four claims went through it on the flagship brief.

The standing rule on all of it, set at the start, is that the outputs are recommendations or a statement that more information is needed. The people who own the decision own the decision. Gaps get named, not written around.