Abstraction-Indexed Canonicity: Bidirectional Compression, Generative Authority, and Meta-Axis Governance

Abstract

Long-lived AI systems may preserve canonical structures at more than one level of abstraction. A concrete data-level canon can coexist with metadata that organizes relations among records and with higher-order generative constraints that determine how schemas, mappings, and lower-level realizations are produced. This paper develops an abstraction-indexed canonicity architecture in which data, metadata, and meta-metadata have distinct canonical standing. Downward expansion and upward compression are both candidate-producing transformations rather than automatic transfers of canonical authority: Expansion ≠ Canonicalization and Compression ≠ Canonicalization. Upward abstraction is tested by registered invariants, residual preservation, and round-trip adequacy; lower-level counterexamples retain an authorized path to reopen higher-level canon. High-generative-reach expansion is separately gated by capability, conditioning context, and current authority because compact higher-level objects can have a large downstream blast radius. The paper then permits multiple generative centers at the meta-metadata level and introduces a separately governed meta-axis that records applicability, compatibility, translation, conflict, and externally authorized selection without collapsing plural centers into one super-truth. The meta-axis is itself scoped, versioned, contestable, replaceable, and non-self-authorizing. The contribution is not metamodeling, multi-view modeling, or model transformation as such, but the canonical and authority semantics placed around compression, expansion, plural generative centers, and their coordination. Prospective predictions, rival hypotheses, and a staged testbed are specified; 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 This paper derives from the author's broader Information Ecosystem Theory, which remains unpublished at the time of this release.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 a companion to Canonicity Without a Truth Monopolist: A Custodian Architecture for Long-Lived AI Ecosystems (https://doi.org/10.5281/zenodo.22711406), which develops the underlying custodian and institutional-memory architecture; this paper is self-contained.