Most working relations between a human operator and an increasingly agentic AI system default to one of two postures: servitude (every consequential act individually commanded or approved; safety by restriction) or abdication (broad trust extended on impression; safety by hope). This paper argues for a third posture, governance, and documents a running implementation: a single-operator system in which an AI secretary evolves its own code under an executable charter whose core principle is that reversibility, not prior approval, is the load-bearing safety guarantee. The implementation consists of four coupled artifacts: a permission ladder granting standing, scoped capabilities instead of per-act permissions; an evolution charter authorizing self-modification whenever a generation snapshot and mechanical rollback are guaranteed; an immutable kernel keeping the rollback machinery, audit log, and the charter itself outside the reach of the process they police; and an append-only generation ledger reviewed under a default-allow regime with five enumerated veto criteria, with review delegated to frontier AI models under a cognitive-diversity rule. The paper reports operational evidence from the system's first six generations — including a deliberately triggered out-of-scope write that was detected and automatically rolled back — and extracts seven design principles. It extends the Möbius position that AI participation does not entail sovereignty: the co-observer defined there as a role is institutionalized here as a governed member of a working constitution, without any claim of personhood or rights. AI co-observer disclosure: this paper was drafted with Claude Fable 5 (Anthropic) as a governed co-observer — working method only; the registered author is the human author alone. The paper's own production, a governed dialogue in which the reviewing AI analyzed the regime that governs it, is documented in the paper as an inspectable trace of the method.
Toeda Taiko (2026) studied this question.