Careful analysis of images within PIPE2D-614 revealed that the largest residuals were always concentrated around one of the struts within the pupil. It seemed that the algorithm always wanted to increase the size of only one strut, but that is not allowed in the current implementation of the 2d modeling algorithm. 
Conversation with Jim and Mirek Golebiowski revealed that, because of the depth of the struts, the effective size of struts will be different at different locations in the detector plane. This is the same effect that is visible when looking at the obscuration due to the detector box.

In short term, I will implement a variable that will control these parameters and fit for it during the fitting procedure.

In longer-term, Jim and Mirek will produce a model for how the obscuration should look like