Adversarial Alternation Under Human Governance

This paper reports a two-day case study in which two AI sessions alternated between builder and auditor roles while one human Governor retained scope, disposition, and ratification authority. The subject was a small verification and evidence-recovery system whose claims could be quarantined when a dependency failed and restored only after re-derivation and human ratification. The case suggests a practical hardening loop: freeze the candidate, replay it cold, attack beyond packaged tests, reproduce every accepted finding before repair, convert each counterexample into a permanent regression, and prevent artifacts from conferring standing on themselves. The source reports that later builder errors were caught by earlier auditor tests and that the last ledger round had no findings under its probe set. Those observations do not establish that alternation outperforms one careful agent. The sessions were role-separated but not shown to be epistemically independent; they were probably from the same model family. There was no baseline, the program was small, the author participated, probe sets changed, and findings were counted in inconsistent units. The frozen archives and full reports were not rerun in this review. The defensible contribution is therefore an evidence-bounded method specification and a set of hypotheses for controlled evaluation, not a causal effect claim.