Failure analysis

Where the model went wrong

Updated

What already works

• The model almost never missed a real line item: 201 of 206 were found in round 1. The errors were extra rows, not missing ones.
• On matched lines, prices were exactly right 99% of the time.
• Customer, document number and currency were right on every sampled document.
• Side-by-side header blocks, a known risk for text extraction, were read correctly.
• The verification script flagged real errors before any human looked: wrong quantities, a total not printed on the page, and a garbled page.