Confidence Scoring Using Whitebox Meta-models with Linear Classifier Probes