Same retrieval evaluation. Different evidence. Different answer.

In the preregistered experiment, answer correctness changed in 17.27% of primary comparisons (equal-weight mean across four datasets) while the full assessed relevance-by-rank pattern and five retrieval metrics were held fixed.