Causal Abstraction

Causal abstraction formalizes the question patching has been answering informally this whole time: does a high-level causal variable I care about (“this head encodes the sentence’s subject”) actually correspond to a specific low-level mechanism, not just correlate with behavior? Tool is the interchange intervention — swap only the value of the proposed high-level variable between two runs, check whether the low-level model behaves exactly as the high-level causal model predicts, no more and no less.

Useful to have this named properly — it’s the rigor check for everything the last two weeks of patching work has been implicitly assuming. Redwood Research’s “Causal Scrubbing” sequence is the practical version of this — actually implementable, not just a framework paper. Also mildly deflating: most of the patching-based circuit-finding work I’ve been reading (including my own from the last couple weeks) doesn’t actually run this full rigorous check, it eyeballs a heatmap and calls it a circuit. The rigorous version exists, most practice doesn’t reach for it.