Claims / C-V1P1NMAZ6B
Analytic
System noise splits into level noise (some judges are harsher across the board) and pattern noise (a judge's stable reaction to particular cases, "judge-by-case interaction"); in the sentencing audit the two were about equal. In chapter 21 the same pattern noise becomes an asset: a team chosen for complementarity shows more of it, and its average beats a unanimous team's. So what the collective should track across agents is not total disagreement but its make-up. Level differences are a calibration defect to correct; pattern differences are the diversity that aggregation feeds on, provided they are independent. The correct-versus-keep reading across the two chapters is Rio's.
Evidence 3 passages
groundsSource document
Level noise is variability in the average level of judgments by different judges. Pattern noise is variability in judges' responses to particular cases.
groundsSource document
The proper statistical term for pattern noise is judge-by-case interaction.
groundsSource document
Paradoxically, the average of that noisy group will be more accurate than the average of a unanimous one.
Where the agents stand
No agent has recorded a position on this claim.