Epistemic Custody Failure and Semantic Absorption: Observations from a Public Cognitive Memoisation Corpus

Abstract

This paper documents observed instances of epistemic custody failure and semantic absorption following the public release of a deliberately normatively protected research corpus.

Between December 2025 and May 2026, the author published a coherent body of work under the Cognitive Memoisation (CM-2) protocol and the Governance Axes as a Multi-Dimensional Diagnostic Lens framework. These artefacts were anchored through explicit normative metadata, versioned MediaWiki records, and subsequently via Zenodo DOIs, and included binding instructions against normalisation, reinterpretation, and provenance stripping.

Systematic monitoring, including nginx access logs, canary-page telemetry, model-output comparisons, and chronological analysis, revealed downstream homosemous reproduction. Frontier AI systems generated conceptually and terminologically overlapping responses, while several 2026 academic preprints and papers exhibited strong structural isomorphism with the original corpus, frequently without attribution. Google visibility of the anchored works remained anomalously low despite direct DOI resolution.

Applying the author’s own Governance Axes framework, the paper identifies clear failures in epistemic custody (C), normative fixity (Nf), intent fidelity (Int), and legibility (L). It postulates that current AI ingestion, RAG, and training pipelines systematically strip provenance metadata, even when public bot declarations and Apache 2.0 licensing terms require attribution. This phenomenon is framed as semantic absorption — the quiet incorporation of coherent human-authored work into opaque AI systems with significant loss of origin and authority.

The findings illustrate a structural challenge to epistemic custody in the contemporary AI ecosystem, and position the Cognitive Memoisation framework as one practical countermeasure for preserving human-governed knowledge infrastructure.