AccuracyReportingLearner should wrap a learner, and try new training examples as predictions first.  when the prediction is provided, then it adds the example.

it should log stats for the learner + task via UMA.  precision, recall (or TP / TN / FP / FN) , maybe entropy, etc.