Ovrin’s confidence model is built from multiple named signals. That is different from a single opaque model score.
Typical signals
The upstream docs describe signals such as:
- grounding
- OCR agreement
- schema validity
- format compliance
- cross-field agreement
Why this is important
It lets you inspect the answer instead of merely trusting a score. If a value is high confidence because it matches the schema but low confidence because it is not grounded, the system can expose that fact.