Beyond Paper-Centric Science: DGP Multiverse Science as an AI-Native Architecture for Scientific Knowledge

Science seeks increasingly adequate accounts of how reality generates the phenomena we observe. Theoretical inquiry proposes, compresses, and relates generative structure; empirical inquiry constrains it through observation, measurement, and intervention. We argue that the familiar focal-hypothesis and paper-centric regime is a historically effective compression under severe limits on human attention, computation, formalization, comparison, verification, and coordination. Artificial intelligence changes this cost structure by making it feasible to maintain and interrogate far larger populations of candidate explanations. We call the resulting paradigm DGP Multiverse Science. Its primary shared object is the Living DGP State: a plural, persistent, and versioned population of explicit candidate data-generating processes (DGPs), rather than a corpus of papers, isolated claims, or one favored theory. We develop the DGP Multiverse Architecture, comprising a Living Knowledge Base that stores this state, a Research Engine that selects theoretical or world-query actions expected to improve or discriminate it, and an independent Verification Engine that authorizes only bounded epistemic updates. These components operate through a Multiverse Revision Loop that maintains, discriminates, verifies, and revises the candidate population. A minimal formal demonstration shows why arbitrarily precise observational analysis can remain non-discriminating between observationally equivalent DGPs while a targeted intervention separates them. Papers remain essential as citable update proposals, but need not serve as the scientific state itself.