What a confidence score actually promises

A confidence score looks like a fact. A field comes back as total: 12,480.00, confidence: 0.99 and it is tempting to read the 0.99 as a measurement. It is a promise, and like any promise it can be kept or broken.

The promise

If a workflow produces a thousand fields with a score near 0.90, about 900 of them should be correct. That property is called calibration. A model can be very accurate and badly calibrated, confidently wrong on a few things, and a badly calibrated score is worse than none because it teaches people to trust it.

How we check it

Every correction a reviewer makes is a labelled example. Nightly, we group a workflow’s fields by score band, compare the share that reviewers confirmed with the score, and adjust the mapping so that the two agree. The calibration chart in a workflow’s settings shows the result: a diagonal line is a kept promise, and a curve under it is a promise we are breaking.

New workflows start with little data, so they start conservative. Scores are shrunk toward lower values until a few hundred reviewed fields exist, which means a new workflow sends more to review at first and less every week.

Thresholds belong to you

We pick sensible defaults, but the threshold is a business decision. A wrong supplier name is an annoyance. A wrong bank account is a loss. Tessera lets you set a threshold per field, and for some fields, bank details being the obvious one, a rule that holds regardless of score.

What a score cannot tell you

A score reflects how well the system read the page. It says nothing about whether the page itself is right. An invoice with a perfectly legible, perfectly fraudulent account number gets a perfect score. That is why we pair scores with checks against your own data: vendor master, purchase orders, previous documents.

Scores tell you when to look. Checks tell you what to look for.