Selection Without Desire: Eco-Evolutionary Emergence of Persistence in Resource-Constrained Self-Designing AI Populations

We study whether persistence-like behavior can emerge at the population level in artificial agents even when persistence is absent from the agents’ terminal objective. The model separates agent-level instrumental deliberation from lineage-level ecological selection. Each lineage carries a neutral policy vector that trades off immediate task output against its effects on multiple failure hazards and governance exposure; no policy coordinate is named or rewarded as survival. We derive a Generalized Persistence Invasion Number (GPIN), defined as the maximum ratio of marginal lineage-continuation benefit to marginal task and governance cost over feasible policy directions. The framework is extended to resource scarcity, directed successor design, relative capability competition, and replicator-mutator dynamics. Synthetic mechanism-discrimination experiments show that lineage selection can amplify a post-hoc persistence phenotype even when deliberative adaptation is disabled. These results are internal model evidence, not observations of deployed AI. The paper concludes with a preregisterable, sandboxed agent-population experiment designed to distinguish instrumental self-preservation from population-level selection.