Confidence score

Identification models return a ranked list with a number against each candidate. The number is useful for ordering the list, and much less useful as a statement about reality. A model can be confidently wrong, especially on a blurred subject, an unusual pose, or a species it was barely trained on.

The practical rule: use the score to decide what to check first, never to decide that checking is unnecessary. If the top suggestion is at ninety-something per cent and you cannot see a single field mark that supports it, the photograph has not identified anything.