This paper proposes an architecture for preserving authority through computational composition. Three primitive types (functional, state, and credential units) compose under eight protocol mechanisms that admit governance, refuse non-reconcilable composition, and witness every act on an append-only ledger. Authority chains terminate at constitutional source credentials anchored at natural persons. The same compositional vocabulary operates at every scale, from a single functional unit through institutional and sovereign substrates through cooperative arrangements between sovereigns. The structural fault the architecture addresses predates the current AI moment. Institutions that run consequential computational systems are the only parties with operational access to them; counterparties, regulators, and individuals depend on the operating institution's cooperation for any visibility into what the system did. This arrangement was tolerable at institutional time-scales. It is not tolerable at machine speed. Twelve cases from public record (including Horizon, Robodebt, Lavender, SolarWinds, the 737 MAX, LIBOR, and the London Whale) are used to show how the architecture's machinery would have surfaced or constrained the failure in each. A four-pattern adoption space across granularity, archive hosting, operator scope, and policy coverage admits incremental adoption by individual operators and by cooperative substrates between them. The architecture's operational property is recoverability of authority through composition rather than prevention of compromise. Its boundedness is named through twelve irreducible failures. Its further contribution is structural preservation of the epistemic content of authority: every act records the calibration under which units asserted their outputs and the policy verdict on that calibration's adequacy. The technical primitives are not novel; the integration under the protocol's structural commitments is.
S Wheeler (Fri,) studied this question.