It's not news that AI hallucinates. In extraction: the field isn't in the document, the model makes a value up, and the API returns it as clean JSON. Nothing errors, the wrong number just lands in your database.
We measured how often, on six APIs including ours.
velrim.com/research/fab...
velrim.com
Extraction APIs invented values for 17% of the fields that aren't in the document. Ours included. — Velrim
Velrim ran this comparison, and we sell one of the six APIs in the table. Every raw output, request ID, and the scoring CLI are in the repo.