Thread:FixSRwiki/@comment-46261745-20200903015444/@comment-26544499-20200910164556

Pursuit Battle should be cut and dry, it's all very near the top of 合戦／勝利処理.

For the losses things, my suggestion is to keep it simple. You could maybe do something to make things easier on lower difficulties, but I see no reason to keep a feature that lets the player exploit the AI or that has so much RNG that the player could be confused or enraged when the AI rolls well. If I were to change it, then in the name of consistency, my first idea would be this: How's that? I've not put much thought in to how it could be exploited, but my gut instinct screams that losing a province is a very big deal, so it's costly to exploit. I can also totally see your point about the Demon Army, but does this cover it?
 * On the lowest difficulty (AI difficulty or star setting?), probably keep things as they are.
 * On the next difficulty, put the entire AI on the same restriction as the player. In other words, if any AI conquers a province, all of the them stop doing normal attacks.
 * On the penultimate difficulty, each individual AI stops attacking when it conquers a province, but the others don't care about another AI stopping.
 * On the highest difficulty, you losing battles has no effect on the AI's behaviour. The mechanic is gone completely.

I'm still not quite sure about how the AI troop sizes work. As I've shown in a webm before, Weaken Enemy Countries confused me enough that I'm no longer sure how they work.

I can remember Somehani doing some coding, but I can't remember what. The massive table was my work. He was focused on picking apart the algorithm for what the AI does during its turn. Speaking of, I've often got some doubts about my results. For example, trivial back-of-the-envelope stats says that the real numbers for that silly 92% thing in the speed algorithm is actually 92.6% (in R, I think it's as simple as "1-pbinom(0,30,1/12)" for the probability that any given unit gets the +10), so I've honestly got no idea where I got 92% from (rather than 93%). I'll shot myself if I find out that I simulated it or made a rounding error, but for now I'm hoping that I knew something about the algorithm back then that I've forgotten now.