Self‑improving policies constantly expose new failure patterns, forcing their judges to evolve in order to provide reliable verification. Fixed judges limit the quality of optimization feedback and hinder the discovery of useful training examples, thereby restricting further self‑improvement. This issue is especially acute in embodied reasoning, where evaluation must account for spatial grounding, causal inference, and safety‑aware decisions.\ \ We introduce VeriFine, a framework that scales verification by co‑evolving the policy, training curriculum, and judge. The system operates through two intertwined loops.\ \ The Policy Improvement Loop employs a rubric‑based judge to diagnose recurring errors, construct an adaptive curriculum, and optimize the policy. When progress stalls and verification becomes a bottleneck, the Judge Improvement Loop activates. It selects informative failure cases for human guidance and refines the judge via co‑active calibration, where humans and agents resolve disagreements and converge toward an objective physical‑reasoning rubric. The updated judge then drives the next round of data selection and policy optimization.\ \ Experiments on driving and robot navigation tasks, using both reinforcement learning and supervised fine‑tuning, demonstrate continuous improvement of both policy and judge capabilities. Scaling verification thus enables ongoing self‑improvement as failure patterns evolve.\ \ Review