Recently proposed bert-based evaluation metrics perform well on standard evaluation benchmarks but are vulnerable to adversarial attacks, e.g., relating to factuality errors. We argue that this stems (in part) fr