Maybe we are quibbling over what "correct" means. If you regrid data and change the reference for sea level, you will invariably end up with z values that don't exactly match the original data set. Going to a new geoid involves some long range, substantial corrections, and some lower amplitude, shorter-range corrections.
For heuristic purposes, try a 2D example. Suppose the original grid has z to just 1 meter, and looks like this:
^
|z
4o
3ooo
2oooooo
1ooooooooooooo
123456789ABCD ->x
...and you must regrid to an x-value that is offset by 0.01. Linear interpolation among the z values will invariably make the new values non-integral -- in fact they will be off-integral by small fractions of a meter. That's just the grid "correction," not a geoid "correction." I guess it might be better to call these adjustments; no one is saying the original data were "wrong."
This article is from 2006:
http://www.ngs.noaa.gov/CORS/Articles/OPUSMexico.pdfThe article talks about constant improvements in the geoid, but I have no idea how one would track what the reference status was when the data were transferred to noaa/usgs. Around here in NV, the geoid change (since surveying) caused shifts of about +1 meter or more for benchmarks -- but there are some areas that "lost" elevation.