Backbone Identifiability in Relativity of Causal Knowledge

A recent paper on arXiv (2608.10664v1) explores the Relativity of Causal Knowledge (RCK), a framework that elucidates how agents with distinct structural causal models share causal insights through a common interventionally consistent abstraction, referred to as a backbone. The authors tackle a previous identification issue: under what conditions is the backbone shaped by agents' private causal knowledge? In a fundamental scenario involving two agents and a common effect, two private causes affect one shared result, with each agent recognizing only the causal marginal pertinent to their viewpoint. The study reveals that, given standard assumptions of compatibility, non-degeneracy, and local overlap, these local causal marginals fail to pinpoint a unique backbone. Countless joint intervention kernels can produce identical private reports while differing on joint interventions. Nonetheless, a conditional recovery finding indicates that additive separability eliminates the hidden interaction degree of freedom, thus allowing for backbone identification. This paper is newly classified and accessible on arXiv.