Beckers, S., Eberhardt, F. & Halpern, J. Y. (2020). Approximate Causal Abstractions. Proceedings of the 35th Uncertainty in Artificial Intelligence Conference, PMLR 115: 606–615. Preprint: arXiv:1906.11583 (27 June 2019), as “Approximate Causal Abstraction”.
Read in full (NOTE-1045).
Why this is here
A25’s first reading: “What, exactly, is novel about governance-response sufficiency?” A25: “We should not present governance-preserving coarse-graining as a new general theory of causal abstraction.” A26 §1 set it against TERM-023, CLAIM-350 and THEORY-820. Bears on CLAIM-407, CLAIM-409 and TERM-065. The record holds causal abstraction only through interchange interventions (THEORY-243); LIT-038’s status note sends a reader wanting causal abstraction to Beckers–Halpern.
Note of 2026-10-27: causal abstraction, now held further
“Why this is here” says the record held causal abstraction only through interchange interventions (THEORY-243). The circuits works read on 2026-10-26 at the owner’s request add two holdings. Geiger et al.’s distributed alignment search (LIT-1551, NOTE-1192) grounds interchange-intervention accuracy in their own approximate causal abstraction, which holds on average over inputs, a cousin of this paper’s approximation. Causal scrubbing (LIT-1549, NOTE-1168), which Geiger et al. count as causal abstraction analysis, tests a hypothesis by resampling and scores it through the partitions of the data it induces (THEORY-978). Status and summary are unchanged.