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

Abstract

Long-lived AI ecosystems need to preserve what was institutionally recognized, when, and under which authority, even after correction. This paper proposes a Custodian architecture for scoped canonicity without assigning a model ownership of truth. It separates record standing, contestation, handling, and attestation; preserves plural canonical branches; and assigns unsupported singleton selection to the query rather than to the records. A query-relative evidence contract distinguishes inclusion, absence, source coverage, freshness, and reproducible historical views. State-dependent approvals bind verification to finalization. Succession retains pending candidates and separates response diversity from authorized transfer; restoration preserves alternatives without reviving historical authority. A generalized event cut covers proofs assembled across domains, while actual detection additionally requires access, comparison, and reporting. The paper states elementary conditional accountability, replay, evidence-survival, and delay results, and proposes comparisons against capable governed-ledger, CRDT, archival, authorization and provenance-analysis systems. The contribution is an explicit canonical application contract. Seven conditional predictions separate conformance, additional information and practical benefit at equal information, with three shared test designs and measurable null conditions. No implementation results or public preregistration are claimed. Note on Version 2.0: this version revises the registered v1.0 (about 14,000 to 17,000 words, 32 pages). It adds a conditional accountability limit, query-relative evidence closure and replay, binding of verification to finalization, attestation grades and evidence coverage, an external-evidence section with authorization freshness, local verification, behavioral diversity and lineage rebuilding, and shared test designs with measurement conditions. Files: the v2.0 manuscript and a supplementary archive containing the review note, a change record, a verification script, a validation record, a manifest and SHA-256 checksums. This paper derives from the author's broader Information Ecosystem Theory, which remains unpublished at the time of this release; it is written to stand alone.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.