[dannyz]

It would be interesting to see some example questions and answers. Since the test is multiple choice is it possible that the model has gotten very good at estimating how likely a possible answer is?