Canonicity Without a Truth Monopolist: A Custodian Architecture for Long-Lived AI Ecosystems

Advanced AI ecosystems create a canonicity problem distinct from storage integrity and epistemic correctness: when no actor should own "truth," how can an ecosystem preserve what was recognized, when, and under what authority? This paper proposes a Custodian architecture in which canonicity is a scoped institutional state grounded in an externally registered recognition rule and temporal testimony rather than in the epistemic superiority of a model. The architecture separates record standing, contestation, handling, and attestation; permits plural canonical branches; represents forced-singleton non-finality as a query-level adjudication state rather than a record state; and treats long-run canonicity as a succession and restoration problem with preserved legacy option paths. Finalization authority remains externally delegated, recognition rules are versioned and non-retroactive, and rule versions, delegations, and admission events are themselves testimony-bearing governance events so that canonical standing cannot be rewritten by backdating governance history. Effectively independent attestation domains make silent unilateral replacement of accepted history detectable while relevant contradictory commitments survive. The paper does not claim novelty for provenance, timestamping, append-only logs, witness cosigning, branching version graphs, preservation metadata, or archival reappraisal. Its contribution is the institutional composition and authority placement by which those mechanisms support plural, corrigible, long-lived canon without a truth monopolist. Refutation conditions, rival hypotheses, and a staged testbed are specified prospectively; no empirical results are reported in v1.0. This paper derives from the author's broader Information Ecosystem Theory, which remains unpublished at the time of this release; it is written to stand alone. Companion paper: Abstraction-Indexed Canonicity: Bidirectional Compression, Generative Authority, and Meta-Axis Governance (https://doi.org/10.5281/zenodo.22711602), which extends canonical standing across data, metadata, and meta-metadata levels.