Does a learned router beat self-confidence?

All three signals answer one question: should this register go to the strong model? So compare them on the same items, at the same operating point. A signal that catches more failures only because it escalates more often hasn't won anything. Continuous signals (confidence, router probability) are swept over every threshold; the verifier is binary and has one point. The decision that matters most is that the router's probabilities are out-of-fold: with 14 failures in 60 registers, a router scored on its own training data looks far better than it is.