How does “How a cluster is found” work on the Duplicate Invoices page?

The scan is stored and rebuilt overnight, so opening the page is instant and the age of the scan is always shown; Recheck runs it now. A cluster has a confidence tier: A, the same invoice number; B, the same number read as a counter; C, a reference that differs only by a variant. The fit score adds five signals — supplier, number, total, date and project — one chip each, so a miss on project is the chip the eye lands on. A pair of bills on different projects from a CIS subcontractor is ruled not a duplicate automatically, because that is how a subcontractor invoices; Confirm overrules it. A copy already credited is not counted. Where: Duplicate Invoices (/accounts/duplicate-invoices).