Asked whether it is sure, Claude changes its answer. What has that change actually demonstrated?

Why this is the answer

A challenge in the prompt is itself an input, and the model can revise a correct answer or defend a wrong one — which is why grounding the answer in a source, rather than re-interrogating the model, is the recommended way to reduce fabrication. Revision does not identify which answer was wrong, no stricter review pass is triggered by the question, and there is no confidence level being disclosed.