The ranker evaluates candidate items by running multiple n-sample median re-ranking passes over full competitor sets and extracting qualitative rationale strings that explain why each item placed above or below the target.