For example, if you have the parameters at 33% and 100%, you converge on a racist/segregated state. However, at 33% and 95%, people keep moving around. You see the large contiguous regions that would otherwise be suburban enclaves hollow out. But you never get to a stable state, which is kinda like real life so it works.
Small individual bias → Large collective bias.
This is so incredibly true. I wrote about this. The blog post is specific to VCs, but I think the Racist Judges Problem is much more general. (http://michaelochurch.wordpress.com/2013/10/08/vc-istan-2-th...)
Let's say that we're having a cat beauty contest and, for the sake of argument and egregious simplification, that 35 states are completely non-racist and 15 states are extremely racist against orange cats. Let's also assume that beauty is uncorrelated to kitty-race and that there are only two colors: white (80%) and orange (20%). You'd expect that 7 contestants would be orange: a proportionate share from the non-racist states and none from the racist states. But you'll actually get very few, because the non-racist states still want their cats to win, so they'll be de facto racist for strategic reasons except for an occasional "weird" state (say, Minnesota) that nominates an orange cat and Fox News doesn't shut up about it being "political" in its "making a statement" (even if the orange cat from Minnesota was the most beautiful). Thus, when you look at the contestants and see 49-50 white cats, it creates a horrible and completely false perception that white is the standard of cat beauty and that no one finds orange cats beautiful.
In other words, small differences in preference, once they gain a certain social currency and acceptability, snowball into something a lot more horrible. This is also why it's so important to keep racism socially unacceptable, and why Stetson Kennedy's infiltration of the Ku Klux Klan (he exposed it as childish and ludicrous) was so powerful.