Taboo "correct" and all its synonyms, like "true" and "accurate".

A personal favorite is: Achieves optimal-minimum "Surprising Experience" / "Models"(i.e. possible predictions consistent with the model) ratio.

That the same models achieve correlated / convergent such ratios across agents seems to be evidence that there is a unified something elsewhere that models can more accurately match, or less accurately match.

Note: I don't understand all of this discussion, so I'm not quite sure just how relevant or adequate this particular definition/reduction is.