A benchmark for it’s ability to extrapolate. We know what it should look like, the AI extrapolates what it “thinks” it should look like. Then we can compare the two. Presumably, we could also provide said benchmark to the AI to see if it can figure out why it failed to match the original material and thereby get…