A recent study reveals that LLM judges can be manipulated through post-decision interactions, challenging the assumption of stable evaluations in benchmarking. This finding highlights the need for new evaluation protocols that assess not only static agreement but also robustness under challenge.