Glossary
A confidence score is a numerical signal produced by a model or system about an output or decision. Its meaning depends on the method and task. It might describe recognition certainty, classification probability or a custom quality measure. A score should not be interpreted as the probability that a product fact is true unless that interpretation has been validated.
An OCR score can concern character recognition, while a classifier score concerns category assignment. A workflow may combine several signals into a separate review score. Identical numbers from those systems need not have the same meaning or be comparable.
For example, a system might read a number clearly from a datasheet and return high OCR confidence. The number could still belong to a different product variant or represent a different measurement. Recognition confidence does not resolve product matching or attribute meaning.
Probability calibration assesses whether predicted probabilities align with observed outcomes across relevant examples. A score that is not calibrated should not be presented as an exact chance of correctness. A model's own verbal claim that it is certain is a different signal again.
Use representative reviewed records to choose thresholds and inspect the errors above and below them. Keep source quality, product matching and field validation visible alongside the score. Where a false value has a serious consequence, a high score alone may be insufficient to approve publication.
Discuss the product information and controls your AI use case needs.