Write a PREreview

Learning systems change internally all the time, yet it remains unclear when such change constitutes an organization: when one comes into being, whether two differently built systems carry the same one, and what remains when it dies. Internal structure is usually compared by representational similarity, by what a probe can decode, or by alignment to a hypothesized algorithm 1, 2, 3, 4. Here we identify internal organization by what admissible internal interventions do to an independently named future consequence; the resulting Consequence Organization, a construction closely related to causal abstraction, is endogenous when the system's own consequence history forms it. We show that, across naturalistic and controlled systems, such organizations formed under stable but not shuffled consequence, persisted and reorganized, could be erased, restored and transplanted, and could lose their causal role while carrier-associated readouts or residual carrier content remained. Their identity transferred across independently trained seeds and, in most directions, across architectures, but failed across an intentionally incompatible cross-consequence family even where internal operators remained nearly identical. In the tested systems, internal organization thus has a measurable lifecycle and a consequence-relative identity that neither decodability nor internal-operator similarity alone tracks.