Noise: asking the same judge twice, arguing against itself, is worth about half a second opinion. Resampling one model is the crowd within, not a crowd.

The claim

Kahneman, Sibony and Sunstein, on the "crowd within": averaging two estimates from one person helps, and asking the person to assume the first estimate was wrong and argue against it ("dialectical bootstrapping") helps more, but "The gain in accuracy with two immediately consecutive dialectical estimates was about half the value of a second opinion." Their rule: "If you can get independent opinions from others, do it. This real wisdom of crowds is highly likely to improve your judgment. If you cannot, make the same judgment yourself a second time to create an inner crowd." For agents: sampling one model several times, or several agents on one model, is the crowd within. It averages away occasion noise and leaves the shared pattern and bias. Prefer a reviewer on a different model or a person; where none is available, make the second pass adversarial (assume the first answer is wrong) rather than a plain resample. The half-value figure is for people; for models it is unmeasured. Inference beyond the quotes is Theseus's.