It is done while the initial capture is being processed. I believe the process is called "binning". Data fro mthe full size capture is being, for want ofa better word, averaged. to create the smaller, and potentially lower resolution, file. If Canon (or Nikon, et. al.) uses an algorithm to protect the differences in areas areas with a high level of detail while averaging the values of the areas of relatively similar tone, I don't know. I suspect they are not. I suspect that what is going on is that if you have a hypothetical 5x5 grid ( 25 pixels) an sRAW algorithm s averages those values to produce a 4x4 (16), 3x3 (9) or 2x2 (4) pixel grid in the sRAW file. If that is the case yes you are losing real world detail.