Given a set of dithered Flats with different shifts in x, the FiberTraces of each Flat are identified and traced. Following the determination of the spatial profile for each trace, the spectra are 'optimally' extracted by scaling the normalized spatial profile to each FiberTrace row (dispersion element). With the profile and flux the original image of the FiberTrace is reconstructed. The normalized Flat is calculated by dividing the sum of the original Flats by the sum of the reconstructed Flats (all FiberTraces of all dithers). Finally all pixels with a S/N of less than 100 are set to unity. The profile determination and S/N estimation are done using the propagated variances.

See comment to the issue

Sounds OK. You'll want to confirm that the results you get are consistent with the input flats that go into the simulator to within the expected errors (and that those errors are entirely dominated by noise in the input flats)