Confidence score
When an output carries real uncertainty, surface it visibly and consistently, a 60% confident answer should never look identical to a 99% confident one. Language, visual weight, or an explicit score can all do this.
Relevant principles
What the research says
Well-calibrated confidence promotes appropriate trust and reliance and shapes how accurate users judge the model to be; mis-calibrated (over-confident) scores erode trust. 1
2
Interactive demo
Expenses · new claim
ScanningReading receipt…