Confidence calibration
Do high-confidence AI extractions get accepted by human reviewers? Bands group reviewed entities by model confidence score.
Entities reviewed
6
Overall acceptance
83%
Well-calibrated?
Yes
Unscored decisions
0
Acceptance rate by confidence band
2 bands with data| Confidence band | Decided | Accepted | Rejected | Acceptance rate |
|---|---|---|---|---|
| < 80 % | 1 | 0 | 1 | 0% |
| 80 – 90 % | 0 | 0 | 0 | — |
| 90 – 95 % | 0 | 0 | 0 | — |
| 95 – 100 % | 5 | 5 | 0 | 100% |
| All reviewed | 6 | 5 | 1 | 83% |
Well-calibrated means acceptance rates are non-decreasing across confidence bands — i.e. higher model confidence predicts higher reviewer acceptance. Currently: non-decreasing (well-calibrated).