From Capability to Endogenous Regulation: Evidence and Experimental Criteria for Artificial Life-Like Organization in AI Agent Populations

AI-agent populations exhibit a growing range of social and lifecycle-relevant capabilities, but a catalogue of behaviors does not identify the processes that sustain a population. This paper asks a narrower question: under a declared system boundary, can maintenance, successor generation, heritable variation, differential propagation, and functional continuity operate as one recurrent process without an external lifecycle controller? It retains an 18-construct taxonomy while separating demonstrated capability, the location of regulatory decisions, policy provenance, and policy revisability. A targeted reassessment corrects the preceding manuscript's treatment of TerraLingua and avoids treating researcher-defined environmental rules as evidence against endogeneity. The reviewed systems provide different kinds of support: population organization, engineered agent evolution, replication components, and candidate ecological integration. Their evidential limits must be evaluated transition by transition rather than assigned from the mere presence of a simulator. A revised sandbox protocol combines controller comparisons with feedback interventions, independent lineage and cultural-inheritance manipulations, within-system interaction ablations, and separately interpreted alternative architectures. Essential lifecycle processes must be demonstrated in the same replicate populations; evidence for adaptive propagation requires more than unequal descendant counts. Transition-specific dependencies replace a compensatory life-like score. Positive results would support a bounded organizational claim, not physical self-sufficiency, whole-population organismhood, open-ended evolution, or consciousness. The contribution is a testable causal framework and a source-audited research prospectus. No new agent experiment or independent replication is reported.