Mechanistic Interpretability aims to produce statements like: "Model M solves task T by doing X."
To do so, many causal manipulations are performed to validate an explanation. But what if (many) other, incompatible explanations also pass the causal tests?
