Built a decision-reasoning engine (Orlog) and wanted to fine-tune a local model for it instead of paying per-call forever. The method (DV-DPO): Run a 3-voice council on each question, produce a synthesis Cross-examine: losing voices challenge the synthesis If synthesis gets revised → DPO pair (chosen=post-revision, rejected=pre-revision) If synthesis holds → no pair (good reasoning produces nothing to learn from) Only genuine revisions under adversarial pressure become training...