Agreement Counts Once

The third of four entries from one conversation on 6 September 2026. It follows Where the Compression Was Trained, which says when to distrust each side; this one says what to do when they meet. The first was The Picture the Maths Was Invented to Replace, and the last is The Parable and the Counterexample. A fifth, The Children's Talk, followed later the same day.

The previous entry only says when to distrust intuition and when to distrust the analysis, and not what to do at the moment they meet. What happens when intuition and logic agree, and what happens when one has to override the other?

Agreement

Agreement is the comfortable case and the least informative one. Two sources agreeing is evidence only if they are independent, and they usually are not. Either the logic was built to justify the gut, which is motivated reasoning in a suit, or the intuition was itself formed by earlier analysis, which is what expert intuition is. In both cases the agreement counts once and not twice.

And the two can agree and both be wrong, when both were trained on the same distribution. Intuition and Newtonian mechanics agreed for two centuries. Flat space felt like the only possible space to everyone until the 1820s and 1830s, when Gauss, Lobachevsky and Bolyai each found that a consistent geometry could be built without it. So agreement is when you act, and when you stop paying for more analysis. It is not confirmation of anything beyond the region both were trained in.

Logic overriding intuition

This is the celebrated case — the Monty Hall problem, the birthday problem, the relativity of simultaneity — and the rules for letting logic win are stricter than the folklore suggests.

• Check the premises before the intuition. A valid argument from a wrong premise is where most counterintuitive results actually come from. Until the premises and the domain have been checked, a surprising conclusion is better evidence of a modelling error than of a broken gut, and the prior on I set this up wrong should be high.
• Then ask whether the intuition is in its trained domain. If it is not — probability and scale being the usual offenders — logic wins by default and the intuition has no standing. If it is, and the expert is seasoned in a kind environment, the disagreement is a signal to re-examine the model, because the gut may be tracking a variable the model omitted. Klein's best-known case is a fire lieutenant who ordered his crew out of a burning house without being able to say why; the floor they had been standing on collapsed moments later, because the fire was in the basement and not in the kitchen where everyone had placed it. His intuition had registered that the room was too hot and too quiet for the fire it was supposed to contain. The model was fine; it had the wrong fire.
• Weigh the asymmetry. If logic says go and intuition says stop, and delay is cheap, stopping to look costs little and going costs a lot. The override should wait for the check.
• Do not override; update. The right end state is not logic winning while the gut goes on disagreeing. It is working the problem until the intuition catches up. Monty Hall — three doors, a prize behind one, you pick, the host opens another door that he knows is empty, and offers you the switch — feels like an even bet and is not, but play it with a hundred doors, of which the host opens ninety-eight, and the intuition flips on its own. A persistent gap between what you can prove and what you feel is a place where you will make the error again under time pressure, when only the intuition is fast enough to act.

Intuition overriding logic

This is the omitted case, and it is legitimate more often than the folklore admits. In a kind domain with a seasoned expert, the residual feeling that something is wrong is data about a missing variable, and the honest response is to go and find it rather than to defer to the equations. The popular-science failure the first entry opened with is exactly this override done without noticing: the picture wins over the formalism and nobody records that a contest happened.

The same independence test governs the machine version. A person who has seen a model's output and then agrees with it is not a second source; the agreement counts once.

And for the writing, this cuts cleanly. A character whose intuition and reasoning agree is at rest. The drama is in the override, and the honest character is the one who lets the logic win and then has to live with the gut still disagreeing — not the one whose feelings turn out to have been right all along.