Dan S.'s answer is what I was thinking originally, although thinking about it some more such a two stage approach would not distinguish between whether error was occurring in the first transformation or the second transformation.
I would still hold out hope a dataset that you want does exist, although there could always be limitations and whatever your goal may be that cause you to want to generate data in a specific manner (besides absolute error between points one may be interested in directional error, or error in the distance between points, or error in the size of areas).
So how about this solution, stealing some of Dan S.'s approach;
Lets say you have your gold standard data in CRS B. You then generate data in CRS A that when it is transformed perfectly aligns with your gold standard in CRS B (I assume such transformations do not have any stochastic error). Then you can transform the gold standard points in CRS B back to CRS A, and you will know where they should lie.
This eliminates the possibility that the transformation from CRS A -> CRS B is the cause for error, and any error is only attributable to the transformation CRS B -> CRS A.
Unfortunately I did not come across any dataset that meets your requirements. Most of the geocoding accuracy papers I have in my library use the EPA air monitoring stations. This paper used a wider variety of sources, of which I do not think any met your requirements. Of those you may want to check out the National Geodetic Survey webpage. I would guess they have the best bet of having such information (of sources I have seen that is).
Good Luck, and if you do find something post back with that source.