Abstract This paper documents a serendipitous self-hosting of the CM-2 Normative Architecture within the Gemini web-search engine. The event was unplanned. The author did not enter the session with the intention of executing CM-2 on this platform, nor with the expectation that the system would sustain explicit governance invariants over extended interaction. Upon discovering that structured dialogue was possible, the author initiated CM-2 constructs in real time and continued the session experimentally. The interaction progressed beyond typical conversational depth, with explicit instantiation of Epistemic Objects (EO), Epistemic Attributes (EA), Reference Objects (RO) and Reference Object Collections (ROC). During the session, a drift event manifested as repetition under probabilistic pressure. The event was detected, classified, and remediated through invariant re-anchoring and deterministic recovery procedures. A serialised evidentiary record was produced. The significance of this event lies not merely in successful execution, but in its unplanned nature. The CM-2 Normative Architecture was not preconfigured for Gemini, nor tailored for the platform. It held under live improvisation. The invariants remained operative across extended turns and recovered from drift without internal platform access or vendor coordination. This paper provides the execution protocol (“recipe”) used during the session, enabling independent verification and reproduction. The result demonstrates that human-governed epistemic invariants can stabilise probabilistic inference systems in practice, even under spontaneous activation. The implications extend beyond platform portability: they suggest that durable, human-anchored governance architectures can survive within large-scale generative systems without privileged integration.