Confidence and Correctness Are Indistinguishable to a Prompted Model on Multimodal Geometry

A self-improving model is usually steered by one of two signals, its own confidence or whether its answers are actually correct. Under reinforcement learning this choice matters, since optimizing toward confidence inflates overconfidence and can cost accuracy while ground-truth rewards do not. We ask whether the same gap appears in a lightweight, no-training, in-context loop. Two vision-language models, Claude Haiku and Pixtral-12B, each run the loop twice on diagrambased plane-geometry problems from MathVerse, one arm carrying forward the examples the model was most confident in, the other the examples that were actually correct. Across four rounds the selection signal does not change mathematical reasoning. Accuracy and calibration are statistically indistinguishable between the two arms, and so are confidence on wrong answers and abstention. The contrast with reinforcement learning is the point. A confidence-versus-correctness distinction that DCPO found consequential under RLHF does not transfer to prompt design, where it changes nothing we measured. What varies instead is the model. At nearly identical accuracy, confidence on wrong answers is 0.70 for Claude and 0.94 for Pixtral, and verified-correct feedback never reduces it. Overconfidence on errors is a property of the model, not of the selection signal or the task.