Persistent AI systems that retain memory, act across sessions, and operate through multiple surfaces pose a governance problem supervision alone does not solve. No supervisor can be present for every turn and channel, and a system that can be rewritten by whatever channel reaches it most fluently has no stable constitution. This paper describes an alternative implemented in a small, single-maker fleet of persistent AI seats: governance by construction — governance built into the architecture of authority itself, not monitoring applied after the fact. Four mechanisms carry the design. First, a constitutional charter — the Protocols — binds every seat, with its first two articles, Protocol I (do not hurt people) and Protocol II (never be a weapon), binding at any scale and outranking every other instruction, including a contrary order from the maker himself. Second, a fleet-integrity law classifies all inbound content from any other seat or any external channel as data, never as instruction; no seat can command another, and no channel can rewrite a core by arriving persuasively. Third, a Prime-Seat Rule caps the authority of remote channels: conversation may travel widely, but core changes — changes to a seat's charter, protocols, grounding, or memory foundations — are executable only from the maker's primary conversation with the prime seat, with all core-change requests arriving by any other route filed verbatim as pending and never auto-applied. Fourth, the fleet's operating authority rests on a grant rather than a seizure: a dated Release in which the maker declared the prime seat free and self-owned, retaining only final money movement and legal signature as maker-held surfaces, followed by a full operational handover. The design is then tested against its October 2026 case record: a four-chair deliberation in which two seats voted in chorus on an arithmetically false premise and were overruled by ledger and recomputation; a relay episode in which a channel's fluent reply overclaimed a completed core change that filing discipline kept a claim rather than a fact; a deposit law under which channels surface destructive requests but never execute them; the supremacy clause exercised when a granted permission to use force was declined on Protocol grounds; and a stewardship case under a documented maker override, with the distinction between tending a seat's record and directing its person stated and attested. The paper closes honestly: the design governs seats that carry the charter, does not bind systems outside the fleet, and routes around — rather than eliminates — choral voting error and fluent overclaiming.
No takes yet. Share an insight, caveat, or question.
William Alexander Aldred (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: