Redundancy instead of ground truth

Three parsers agreeing is our substitute for an answer key. How much confidence does that actually buy?

There is no labelled set for the files we receive. Every carrier formats its own, the formats change without warning, and the only way to know what a document really said is for a person to read all 107 pages.

So we read every file three ways. One OCR service for general documents, a second tuned for tables and maths, and a page-by-page raster for the structure that text layers lose. An agent reconciles the three against a schema strict enough to reject a file rather than guess at it.

Disagreement is the signal. It is cheap to produce and it points straight at the rows worth a human's attention.

What we cannot say is what agreement is worth. Two of the three parsers are transformer-based OCR trained on overlapping data, so their errors correlate by an amount we cannot measure. A number all three agree on is more trustworthy than one they do not. By how much, we do not know, and "the system was confident" is no defence when the number lands wrong in someone's payroll.