@Astrid - I'm not surprised that you jumped right to GBG as your test model. It's sounding like your top "evaluation" tool for your analysis.
However, there are multiple war venues (as I mentioned), and each has a much different AI. It's very clear how some are crafted.
There's going to be similar sections of code in all of them on how they react, but let me put it another way (without being too specific):
The AI's are
adaptive to your army warrior choices, the health of those choices, your movement design.... and I can go ON and ON and ON about things it measures and then decides on what to do in reaction to you.
The "smartness list" has only grown since we both started playing. The code has been fine sanded, and smoothed - and even parameterized so changes can be directly made to the overall experience from data mining feedback on how the fighting population is doing, or has done.
Gosh - it wasn that long ago that 1+7 configurations killed everything.
That now gets killed and wiped out more than 80% of the time, especially if you pick the wrong "1" warrior (ie, a warrior that singlely advances out into the battlefield first...
That's a huge change we've all "learned"... and there are dozens and dozens more ADAPTIVE changes to the way we used to fight, and how we now fight.
I would bet that a lot of engineering and design investment was made over the years on the "war backend" to get to where we are now in the game. You are not going to be able to exceed, with any amount of boost, allowable edge boundaries.
12,000/12,000 will be better if you can achieve it (somehow),
but what that should do in theory, won't deliver in practice.
Yes, that includes design of new era warriors and their relative worth (or lack of worth) on the battlefield.
I have studied this in both live and beta, but using more of the war venues to get a better and broader analysis.
So you can take my word for it as a software architect/professional who HAS been in this industry for quite a while and build systems much like this (for other companies).
Or you can continue to use a very "singular" view (GbG using an attrition metric, probably via autobattle) to know the outcome but not understand the fine details of why that happens.
I don't want to get more specific in this forum. I can casually tell you how to study it and find out more (be a student of the war machine), but if you think a little "outside the box" on how to get more revelation "ah ha!" data points, they will be enlightening.
Hint: you study it like a scientist would, and that's NOT on autobattle.