Comment by wuschel

My guess would be that this is a matter of fine-tuning the model to the problem. You would need to run the review experiment N times, observe the distribution of answers, and then fine tune with model operational parameters.

Perhaps someone more knowledgable could jump in here to clarify?