I'm going to make a few miscellaneous factual points. First of all, I don't know of any implementation of the Elo system currently that uses a normal distribution. Elo's original formulation did, though it can actually use *any* curve for the Win Expectancy. And all the big ones now (so far as I'm aware) are using the Logistic curve.
Secondly, Goko doesn't use Elo. Further, we don't know *what* their underlying WE curve is.
Okay, opinions. Goko's system would be much much better if it were the Elo system (with a reasonable choice of K value and underlying WE curve) - of this I am pretty confident. Well, okay, it's really more to do with the choice of WE curve as well as their updating procedure. But I am pretty sure their curve - if they're even using one, which I'm not totally sure they are - is DREADFUL, much worse than a Gaussian would be. And actually, if you look at what they are saying, they have to pervert their system to not allow WE of over 100%/less than 0% in some games, which is prima facie quite bad.
Okay, the 10% sure win but-then-change-that-based-on-how-big-a-rating-gap-there-is is really wanting a different curve. Which is fine, of course, though finding exactly which one is best is EXTREMELY tricky.
You also probably don't want to have automatic cut-offs like that. Let the win% float from 0 to 100, just have a good enough system that (if it's unreasonable to expect a 99% winrate) makes it very difficult for a 99% win-rate to emerge. Which actually isn't all that hard to do, though again, finding the best curve is incredibly difficult.