The instinctive fix for an unreliable model is a second model to check the first one, or a panel of several, on the assumption that independent checkers are unlikely to fail in the same way at the same time. Correlated error is what happens when that assumption is false: the checkers share training data, shared architecture choices, or shared procedure, so they tend to miss the same cases rather than catching each other’s mistakes.
The statistics are unforgiving once correlation is present. For a population of deciders with a positive mean pairwise error correlation, the variance of the aggregate error rate does not shrink toward zero as more checkers are added, it converges toward a floor set by that correlation. Past a certain point, a fourth or fifth reviewer buys almost nothing, no matter how independently they were built or hired.
The fix is not more checkers. It is engineered diversity: different input representations, different data sources, and deterministic checks against an invariant that does not care which model or reviewer produced the answer. A ledger balance either reconciles or it doesn’t, and that check doesn’t get weaker because everyone upstream was confident.