Testing extraction on real shipping files, not a prepared happy path
A test you wrote yourself will pass every time. So we test on the documents forwarders actually receive.
When you write your own test documents, you set the exam and grade it. You will score full marks and learn nothing.
So our test bench runs on real shipping documents from a live desk, with every result compared field by field against what an operator confirmed. Most of the PDFs in that set have no text layer at all. They are scans.
Every change to an extraction prompt is run against the whole set before it ships. If a field that used to be right becomes wrong, the change does not go out.