Abstract

The Human Moral Archive Framework (HMAF) is an external archival infrastructure for the durable preservation of morally relevant characteristics of AI-mediated system behavior. It is not an ethics system, a moral authority, or a mechanism for evaluation, scoring, or control. HMAF exists to address a structural limitation in contemporary AI governance: the inability to reliably examine moral-relevant system behavior across long time horizons as systems, policies, and interpretive standards evolve. The framework establishes a fixed canonical measurement substrate composed of invariant attribute identities, applied externally to observed system behavior and interaction context. These attributes function as measurement units rather than moral claims. Their naming and structural position do not change. When the applied meaning or interpretive guidance associated with any attribute evolves, such changes are declared explicitly through registry epoching and preserved immutably as part of the historical record. This separation between measurement structure and semantic governance enables longitudinal comparison without retrospective reinterpretation. HMAF deliberately excludes embedded evaluative mathematics, normative models, and human judgment from the archival substrate. All analysis, interpretation, and regulatory or ethical assessment occur outside the framework, under explicitly declared assumptions. By separating preservation from judgment, and by externalizing moral-relevant records from adaptive AI systems, HMAF provides a stable evidentiary layer that supports auditability, oversight, and retrospective analysis without constraining system design or delegating moral authority.