Add qualitative performance validation

[wassname]

This helps find out which methods work best, which models it works with, and so on. For example this show that middle layers work best, it only works on model >4B, and diff_pca is best

[thiswillbeyourgithub]

it only works on model >4B

Just in case: have you checked wether the error message Error at layer 27: 'Qwen3DecoderLayer' object has no attribute 'set_control' repeated over and over at the bottom is an issue here?

[wassname]

oh yeah that is just because those layers were not transformed into repeng control blocks. And I just used a try except pattern over all layers

With small models, they are kind of incoherent to start with, and I suspect they don't have well-developed inner concepts, making them harder to steer.