Your point that consensus is not correctness is the key risk: errors from models trained on overlapping data tend to be correlated, so three model families reduce but don't eliminate shared blind spots. Cross-checking is strongest when verifiers use different evidence, such as separate retrieval sources or tool outputs, rather than only different models reading the same context. A useful metric would be how often the reviewing agents change the lead's answer versus rubber-stamp it, which would show whether the extra spend buys real disagreement.