THE
GREAT
CONFLICT

We switched off individual AI components one at a time and checked whether the Axis powers performed worse as a result. In the first pass, it looked like one particular coordination system was essential for everything else to work. On closer inspection, that turned out to be a software bug that had accidentally silenced this system across every test run — not a genuine dependency. We found the bug, fixed it, and reran the test cleanly.

After the fix, we could not demonstrate a clear, independently measurable effect on win rate for any single component tested. That doesn't mean these systems are useless — it means their individual contribution is either small, or only shows up in combination with everything else, in a way our current sample size can't yet cleanly separate. We'd rather say that plainly than claim a positive result we can't actually back up.

Along the way, we uncovered several real, independent bugs we might never have found otherwise — and fixed them. That's the actual point of a study like this: not just confirming what you expect, but surfacing what's actually going wrong. This isn't a one-time exercise: even as we write this, we're systematically working through further, independently found bugs — the AI is under continuous review, not just checked once for this study.

This ablation study alone involved over 2,000 automatically played test games — computer versus computer, so we could measure effects at a scale that would never occur in normal play. A further, continuously growing test run is currently underway in parallel, among other things to check how well the AI plans and executes rare, historically difficult operations like an invasion of Britain (Operation Sea Lion).

Honestly: an AI that plans across multiple layers — look-ahead evaluation, learning components drawn from past games, and coordination between factions — but not an unbeatable or flawless opponent. This study itself shows that: we couldn't demonstrate a clearly measurable strength gain for individual AI components on their own, and the Axis powers structurally struggle to convert an advantage into an actual win — that holds regardless of whether a human or the AI controls them, and is a known balance property we're working on. All figures so far also come from computer-versus-computer games; how that feels against a human opponent may differ. What we can promise: an AI that's tested seriously, with real statistics, and whose weaknesses we'd rather disclose than hide.

That's something we honestly can't answer with confidence yet. Every number in this study comes from computer-versus-computer games — which tells us something about how AI components behave toward each other, but nothing about what a game against the AI actually feels like: whether you're facing an opponent worth taking seriously, or something closer to a rule mechanism working through a checklist. What we can say: architecturally, the AI isn't built on simple if-then rules — it plans ahead, tries to read opponent behavior, and coordinates across factions, which is a different thing from a rigid rulebook. Whether that translates into a "real opponent" feeling at the table is an open, honest question we can't yet answer with data.

Human-versus-AI isn't just something planned for later, either: these games are already a normal way to play today, and that data already feeds into the AI's learning pipeline (weighted lower than pure AI-vs-AI games, so individual human playstyles don't disproportionately skew training). So the AI does learn from games against real players in principle — we just can't yet make a solid claim about how strongly or how quickly that shows up, until enough of these games have been systematically evaluated.