One important caveat is that the non-random strategy would need to be invertible. For e.g. a ml model doing binary classification this is easy (switch the class

One important caveat is that the non-random strategy would need to be invertible. For e.g. a ml model doing binary classification this is easy (switch the classification), for most other things it's not clear what invert means.