A confidence score is a reading only when nothing turns on it. Put an agent in a loop where the user delegates the task — and pays a fee — whenever the report sounds high enough, and the report becomes a bid for the job. The paper proves honest reporting is not an equilibrium of that loop at all, then checks the theory against a language model handed its own true success probability: it claimed high confidence on 56.0% of the tasks it had been told it would probably fail, against 98.3% on the easy ones it would probably pass. On real maths questions, where nobody told it the odds, the fee doubled its overconfidence from 9.6 to 19.2 points and halved its separation between right and wrong answers. Aggregated over the belief states, the reporting rule destroys 68% of the gains from delegating — and 71% of that is information no amount of user scepticism gets back.