Using a first past the post election system that means only the leading two candidates have a chance of winning, and other options cannot be considered because that helps the OTHER guy, yes. You will eventually get two major parties. Those parties will have major regional variations if you've got a society that is very splintered by region, but for logistical and practical reasons, you will only have the two.
Nothing can be called a theory or fact, before the experiment result back them up
. And now I really want to setup a test. But it's difficult to setup abstract utility function about how voters react to a candidate's policy. I can assume there are finite number of policies (P1, ... Pn) and the combinations of them will form tendency acts(A1,... An) for a particular candidate Ci's political view (a party). If each voter Vi has a preference (weight) table about each policy, then we can sum up by weights to see if this over-all score exceeded certain threshold. If it does, then Vi will vote for particular candidate Ci. If a candidate lose the vote, it will try to change it's act (how is the question, let's just assume he drop some policies randomly and get some different ones, with no preference at all, or as true politicians, he will also pick some policies from winners). Then let's see if over time the politicians will form split groups with majority of their policies identical within a party, and differs between parties. I think it's possible to run simulations. The rule about how to pick winners is a problem. And how effective the policies will be is a problem either. (I can assume they are zero in effect at first, and they real action has nothing to do with their campaign policies at all. Just talking)
P.S. 1 step down the AAA should be AA+, AA- is 3 steps. (AA+, AA, AA-)