I have been tinkering with environmental novelty for the past few weeks, and I finally managed to get to what I wanted: a map of expected model performance based on distance to the data used for training. Now to write the paper when I have a moment (like... dec. 2028).

10 likes 1 replies