The origin seems to happen, at least for some of these spots because:

1. The analysis fails because only spots from one side of defocus are analyzed. In that case, we are essentially doing the extrapolation from defocused to focused measurement and that can fail spectacularly. 

I am showing an example of a bad example (Nez5_bad_spot.png) and a nearby good spot (Nez5_nearby_good_spot.png) that illustrates this. 

I have added in my analysis module (internal version 0.24a) a requirement that for the analysis to be successful, there need to be 2 large images from each side of defocus (where the large image is defined as the image taken when the movement of the slit is larger than 1.5 mm )

Additionally, some spots have relatively bad solutions even though there are images available at both sides of the focus. All of these defocused images have areas that are saturated. It turns out that during the reduction of the data, when one uses the option isr.doSaturationInterpolation=False the code does not ``grow'' the mask covering the saturated areas, i.e., it flags the pixels just next to saturated pixels as ok. A large number of such pixels with wrong values that are deemed to be acceptable for my analysis leads to wrong solutions. This has been discussed in #drp-2d channel and I am posting the most relevant part of the conversation as an image here.

Short terms solution is to remove such images from my analysis. Medium-term solution would be to reanalyze and manually grow the mask. Long term solution is for the ISR pipeline to capture this problem.