Finite Regenerative Containment for Adversarial AI Evaluation

Abstract

This preprint presents a finite regenerative containment architecture for bounded adversarial AI evaluation. It brings together the architecture, formal model, empirical program, bounded local findings, and explicit limitations. H1 receives bounded architectural and implementation support in the declared deterministic configurations. H2 receives bounded simulation support under the tested policies and distributions. H3 through H10 remain indeterminate as full hypotheses. Physical, virtual-machine, air-gapped, and real adversarial containment remain untested. The accompanying record includes the empirical paper, a plain-language reader aid, deterministic scientific models, experimental artifacts, tests, source records, and content-addressed verification materials.