In some cases, a model's uncertainty remains stable throughout a reasoning chain and collapses only when the final answer is stated