Provably Reliable Classifier Guidance through Cross-entropy Error Control