A single model reviewing a transaction will reach a conclusion based on its training data, its feature engineering, and its calibration date. That conclusion may be correct. It may also reflect a distributional shift the model has not yet seen, a feature that is correlated with protected class membership, or a confidence score that does not reflect genuine uncertainty.
A second opinion—produced by a different model architecture, trained on a different dataset—provides a structural check on these failure modes. In high-stakes decisions, that check is not optional.