Individuation in Artificial Agents: Environmental Divergence is Necessary, Agency Amplifies

Whether two instances of an artificial system are individuals or merely configurations of one system is usually asserted rather than measured. We introduce a controlled transplant assay that answers it empirically: a component is moved from one instance into another, and the resulting performance change is compared against a self-transplant control. If foreign components cost more than native ones, the components were adapted to their own instance, and the instances are not interchangeable. Applying this assay to a 2×2 design (shared vs. divergent environment × passive perception vs. closed sensorimotor loop; n = 20 replicates per cell), we find that environmental divergence is necessary for individuation — agents in a shared environment score exactly zero regardless of agency — while a closed action loop amplifies individuation by 1.61× when environments differ (d = 0.96, p = .005) and produces none on its own (p = .359). We further show that individuation is localised: transplanting components separately, the two that touch the environment carry all of it (64% predictor, 36% policy) while purely internal components carry exactly zero in every replicate; that it saturates, showing no growth between 200 and 1200 cycles of life (p = .26); and that numerically distinct states need not be individuated at all, since same-world agents whose weights differ substantially score zero because their states remain functionally interchangeable. We report a failed replication, a mechanism that did not work, and a measurement confound discovered and corrected during this work — including a version of the assay that scored +0.081 on bit-identical agents. The assay, the environment, and the analysis are released in full. We make no claim about phenomenal consciousness; individuation is a distinct and measurable property, and it bears on the agency and embodiment indicators in current consciousness-assessment frameworks.