I have seen some significant shifts in predictions recently. Some changes don't appear to be supported by the data. For example you reduced predictions across the board at the University of Minnesota. However, I did not see any denied results for applicants with high (3.8 unweighted GPA and above) and high test scores (30 ACT and above). This sub set of applicants has a 100% admit rate -which is what Parchment predicted previously. Actually, it the Parchment scatter plot shows that everyone with a 3.5 GPA or better and 30 ACT or higher was admitted.
It appears your model(s) lack some interaction factors with regards to unweighted GPA and test score on prediction. Perhaps some additional non linear factors merit consideration.