Is there no confidence score on a “per field” basis?

We’d need to flag any document that returns a non-null value but has a low confidence score.
(cross posted from Slack by admin)

The short answer is that (most of) our methods are deterministic with the exception of OCR-based methods, so we’re not choosing between possible extractions for a field. Rather we either find an extraction or we don’t. This is in contrast to, say, character classification in OCR where you might be trying to guess whether something is an L or an I. That said, I understand that what you’re getting at is how do you determine which docs might need further review.

For that purpose validation (both within Sensible and domain-specific) can play a key role. Check out the validation docs here.