I've looked at the color components in various intermediate stages in gsciemap.c, especially in gx_cie_remap_finish. Whenever I had values of my own to compare with, I found that the ghostscript values were systematically below the ones I had calculated, by something around 0.003. Note that this is near 1/256, so this might be related to incorrect rounding or to a difference in scaling 1.0 to 256 or to 255. I'm not familiar enough with the ghostscript color mapping code to investigate this myself, but if someone wants to investigate it, I could provide examples of systematic differences with respect to "correct" values calculated in double precision.
Assigning to Dan.
Note to self - reference bug #465921 - Retest if changes made.
Note: I have been checking the math in the color look up logic. There seem to be two sources of error. There is no interpolation done with table entries which are less than 0.01 apart. This is controlled by a value in gscie.h. The table entries are clamp to the output range. As a result, there can be errors in the interpolation if one table entry is clamped and the other is not.