The logical steps a model takes between reading input and producing a final answer, which must be correct for reliable reasoning.