Dunno... if you expect guesses to follow a reasonably normal distribution, the mean and median should be the same. I imagine that if you did it with real jelly beans, you'd find that the mean and median of guesses are pretty much equally good estimators of the actual number of beans in the jar; they should be close to identical. The difference is that the mean will include all the wacky far-out guesses and the median won't; so the mean is more susceptible to manipulation. Whether you want that in the real-numbers game depends just depends on what kind of participant behaviour you want.

It'll certainly be interesting to see what kind of actual distribution of guesses you get from this.