Reading a hundred thousand documents without losing the date

A construction record is not a single archive. It is a dozen archives with different conventions: emails with their own timestamps, site diaries dated by shift, minutes dated by meeting and issued weeks later, programmes whose data date differs from their issue date. Treating these as one undifferentiated stream is the fastest way to build a chronology that collapses under scrutiny.

Three dates, not one

Every extracted event carries at least three dates: the date the event occurred, the date the document recording it was created, and the date the document was received or issued. Most disputes about sequence are, on inspection, disputes about which of these dates is being relied upon. Keeping them separate is what allows the question to be answered rather than argued.

Custody as part of the fact

Who held a document, and when, is itself evidence. A revision that existed in the designer’s system on one date and in the contractor’s inbox three weeks later supports a different attribution than one transmitted the same day. Provenance is not metadata attached to the fact. It is part of the fact.