I worked on (wouldn't say I completed) a testing framework for ML models, where you can specify a series of tests for a model to pass. The unique part is emphasizing support from the training data for the inferences you expect the model to make, i.e. checking that similar training data exists for some test cases and is influential in predicting them. It's pretty niche, and I don't explain it very well, but I remain convinced that with the right framing it represents a more rigorous way of making sure machine learning models are built and used "responsibly".