Reader agreement
The vision reader and the OCR + text reader work on their own. When they both return the same value, it is almost always right.
Extraction models are usually confident, but sometimes wrong. Lingua Ledger never relies on a single reading. It collects independent evidence for every field and turns it into a clear confidence score you can trust.
The vision reader and the OCR + text reader work on their own. When they both return the same value, it is almost always right.
We find the value on the page and use the OCR score for those words. This is also how we draw the boxes on the page.
Rule checks: line items add up to the subtotal, tax math is correct, dates make sense and are in order, and IBAN and VAT numbers are valid.
Each reader reports how sure it feels. This is the weakest signal on its own, so it only gives the score a small push.
When the two readers disagree, a vision model looks at a crop of the page to help score confidence. We still keep the merged text because replacing it barely helped accuracy in tests.
Every posted document is saved to ledger memory by vendor. Future documents from that vendor are checked against it for duplicate invoice numbers, changed bank accounts, unexpected currencies, invalid tax rates, or future dates. Your reviews retrain the calibrator, so review work drops as the ledger learns.
All signals are combined into a raw score and calibrated using isotonic regression tested on unseen documents. After that, a 0.9 confidence means the field is right about 90% of the time. The Evidence page shows the reliability curve before and after.
EvidenceAny field below the review threshold goes to a person, along with its box on the page and the evidence behind it. Raise the threshold for higher accuracy, or lower it to reduce human review. The Evidence page shows this trade-off on real data.
Workbench