Load a recording, pick a candidate pair, and run arbitration. Both candidates are evaluated symmetrically — no faster/slower bias. The result is a descriptive classification, not a BPM recommendation.