From the notebook, using DCR templates for the crowded field simulations results in many more false detections in the majority of images. These false detections appear to be caused by artifacts related to overlapping bright sources, and are relatively consistent compared to the false detections using CompareWarp templates. It is interesting to note that there is a dramatic difference in the number of false detections depending on whether the observed field is rising or setting at the time. Most of the observations that went into building the templates were made while the field was rising, so the CompareWarp template performs badly when the science observation is taken while the field is setting.
Based on these differences, I believe the problem with the DCR model was that the shifts of the sub-bands used to build it were too large, and flux from overlapping sources is getting mixed. The fix would have to be to use more observations to build the model, and to use more sub-bands with finer frequency resolution.