We must have a different definition of arbitrary. OP ran 2.3 million tests comparing random battles against the original implementation? Which is probably what you or I would do if we were given this task without an LLM.
I'm very suspicious of such projects so take it for what you will, but I don't have time to debug some toy project so if it was presented as complete but the instructions don't work it's a red flag for the increasingly AI slop internet to me. I'm saying I think they may have used one simple trick called lying.