If ETS is doing their job correctly, scaled scores should be stable across test dates.
To illustrate: let's suppose I got 43 out of 62 questions right on last April's test, which corresponded (I think) to an 800. Theoretically, if I were to have instead taken the October test, because the difficulties of the exams fluctuate, I might have gotten only 35 questions right. However, I should still have scored 800 +/- some error term.
On the other hand, percentiles are not stable. They are revised periodically (once a year?) to account for the fact that test takers are getting progressively better. In April, an 800 corresponded to 83%. But, because more people have been scoring higher, that was revised over the summer, so an 800 now only corresponds to 82%. As I mentioned earlier though, the revision of the percentiles happens to everyone, both old and new test takers.
Thus, getting all the questions right (the raw score) is more impressive on particular test days, but that should be normalised in the scaled 3-digit score.