Research questionHow can alignment systems infer the multiple criteria behind human pairwise preferences?Pairwise preference labels record which option people choose, but not the interacting considerations behind that choice. This makes it difficult to build preference models that are both faithful to judgments and interpretable.