Tagged causal intervention · show all articles
A circuit that looks real on your screen has to survive a matched control, a blind check, and someone else's attempt to break it before it earns a place in an audit report. This is how the practitioners who take that seriously actually work.
A feature-visualization narrative that looks right and a circuit that has been checked against a known answer are different kinds of evidence. Most published claims sit closer to the first than the coverage of them usually admits.
Taking a mechanism apart tells you what each part was holding — but only once the part is off, and only for the mechanism on the bench. What interpretability has established, and what it has not.