The Chris Hirata's simulator (i.e. current ETC) allows a user to choose "sky subtraction error" as a parameter. But this accordingly adds a random error per detector pixel on top of simulated spectrum. Meanwhile, in the reality, sky subtraction error appears as a residual of sky spectrum and the amplitude of the residual varies from exposure to exposure depending on how wrongly the data are processed e.g. in the spectral calibration. 

To estimate the expected quality of a spectrum at a given integration time, it is crucial to simulate a spectrum in such a way as reasonably consistent with the reality. The key uncertainty in doing this is the amplitude of the sky residual at a given exposure, but perhaps a starting point is to assume that it randomly varies from exposure to exposure with a variance equal to a certain percentage (e.g. 0.5%, 1.0%) of sky.

Possible modes of sky subtraction residual are:

• systematic residual with random scale (probably due to fiber throughput variation)
• wavelength shift (due to the wavelength calibration uncertainty, which is expected to be very small)
• variation of PSF profile (tested with arc data)
• spatial variation of sky (see Murdoch thesis)
• variation of dichroic throughput (at transition wavelength)