Per Kirk's suggestion, look into query characteristics that lead to performance increases/decreases with each model.  This should inform query performance predictor implementation.

Document results in Wiki and put plots, if appropriate, into Github.