You’d be quite surprised, I think. Fine tuning a model on one axis can have drastic impacts on another that as a human we would expect to be completely unrelated.

I have never seen anyone argue that this cannot be overcome with more high quality RLVR data.

The practical reality is that the Chinese and American models might have very different politics. But the most relevant factor in model performance is the quality and volume of training data, not ideology of the base model. Unless you are suggesting something very particular about the way Grok was neutered.

Yet empirically it does

It's all a bunch of weights isn't it? Why wouldn't fiddling with some parts of the weights have cascading effects?