As easy as PIE: understanding when pruning causes language models to disagree