A five-lab study finds text color and contrast alone can shift vision-language model verdicts, with Qwen2-VL most affected.