Prompting strategy effects depend on judge model and language in LLM evaluation

A benchmark of 9 strategies across 4 judges and 27 languages shows that the choice of judge dominates score variance, strategy effects reverse across models, and low-resource languages amplify instability.