Interpretability itself becomes a benchmark — passes the burden of proof to → Features proved by intervention

The autointerpretability score is explicitly correlational and says so in its own definition: it is "the correlation between the model's predicted activations and the actual activations" (sparse-autoencoders, §"3.1 Interpretability at Scale", p. 3), and the paper concedes that in its current form the protocol is "unable to verify outputs by looking at changes in output or other data" (sparse-autoencoders, §3.2, p. 5) -- a scorer that can certify a feature looks like a coherent concept but cannot certify that the concept does anything to the model's behavior. Causal proof by patching supplies that missing half, and the handoff is load-bearing at the exact point where scoring weakens: the autointerpretability advantage over baselines "shows minimal improvement in the final layer" (sparse-autoencoders, §3.2, p. 5), yet the IOI patching experiment is run on layer 11 of a 24-layer network (sparse-autoencoders, §4, Fig. 3 caption, p. 5) and still finds dictionary features reaching a given KL divergence in fewer patches and with smaller edit magnitude than PCA. A correlational grade taken alone would have undersold exactly the depth at which the causal test still finds an advantage, which is the clearest evidence in the paper that neither method could stand in for the other: scoring alone would have missed a real late-layer effect, and patching alone, run on every feature without any prior grade to prioritize candidates, would be too expensive to serve as the paper's primary evidence.