google/TrustScore

To Trust Or Not To Trust A Classifier. A measure of uncertainty for any trained (possibly black-box) classifier which is more effective than the classifier's own implied confidence (e.g. softmax probability for a neural network).

To Trust Or Not To Trust A Classifier

This is not an officially supported Google product

Signal for model confidence for a trained classifier, computed based on labeled training examples and the classifier's hard predictions on these examples.

See https://arxiv.org/abs/1805.11783