Peer-to-peer collaborative systems operating offline face a problem that centralized architectures dissolve by fiat: when no single node holds all the data, and peers connect only sporadically through mesh networks, what does consistency even mean? The framework of IntersectionAtomicity and IntersectionCC answers by shifting the unit of reasoning from the global state to the overlap — the specific data subsets that two peers share. Atomicity and consistency guarantees apply only where interests intersect, not across the entire system.
This reframing is more than a relaxation of classical distributed computing constraints. It acknowledges that in a sparsely connected, partially replicated world, demanding global consistency is not just expensive but meaningless — peers that never share data have no consistency relationship to violate. The formal model lets developers reason about what guarantees actually matter for their users, defined by the geometry of data overlap rather than by architectural ideals.
Reader note (2026-11-17): Checked against the full text of arXiv:2603.03899 (Nair, Marino, Pascucci, Brown, Silva, Cummings and Power; v1 of March 2026), which I had earlier held as a weak read. The framework's core matches: IntersectionAtomicity and IntersectionCC as models for peers that subscribe to subsets of the data, with guarantees applying to the data within the intersection of their subscription sets. The paper is a short modelling and guidelines paper: I searched for proof, prove and theorem and found no hits, and for implement, evaluat, experiment and prototype and found only mentions of other systems' CRDT implementations, so the essay's statements that the problem becomes solvable and that the formal model is shown to map to how collaboration works are not results the paper reports; what it reports is the two models, the conditions under which they are upheld, and a set of design guidelines. The paper also identifies a limit the essay does not mention: peers can form data islands when groups of nodes with overlapping interest sets are insufficiently connected, and a single node cannot form one because it can infer that it must receive data from somewhere.
The structural insight: guarantees in distributed systems are not properties of the system as a whole but of the intersections between participants. Consistency is local to shared context. When you stop demanding that the whole world agree and instead ask only that overlapping views cohere, the problem becomes solvable — and the solution maps to how collaboration actually works.
(arXiv:2603.03899)