Confidence and explainability

How Ovrin scores an extraction: six named signals combined into a confidence value you can take apart, and why it ranks results rather than predicting them.

One of Ovrin’s strongest product claims is that confidence is not just a single opaque number. It is broken into named signals and tied to provenance.

Why this design matters

An extracted total is not just "0.98 confidence". The score is composed from six named signals,
each recorded on the field:

  • grounding — whether the value actually appears in the source
  • agreement — whether two independent readings produced the same value
  • ocr — character-recognition confidence over the words behind the value
  • schema — whether the value satisfied its declared type and rules
  • format — whether it matches the expected shape for its kind
  • cross_field — consistency with sibling fields

A signal that does not apply is absent rather than zero, and the remaining weights are redistributed.

Explainability

Each field can point back to where it was found in the document. This supports human review, auditing, and debugging of false positives.

Important caveat

Confidence is a ranking signal, not a guarantee. Upstream Ovrin docs explicitly note that confidence weights are provisional and should not be treated as a probability claim without calibration.