Local Validity Is Not Closed Under Composition: A MoCKA Framework for Composition-Aware System Assurance in Consequential Human–AI Systems

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This external-review preprint presents a bounded theoretical proposition for consequential human–AI systems: local validity of individually evaluated elements does not, in general, establish composition-level admissibility. The paper develops a MoCKA framework for composition-aware system assurance, including a formal composition object, authority and Human Gate distinctions, two-axis epistemic and operational status semantics, and a reproducible benchmark protocol for consequential composition. The central proposition is intentionally limited. The paper does not claim that all compositions are unsafe or inadmissible, nor that local validity is irrelevant. Rather, it establishes that local validity alone is insufficient to derive system-level admissibility for consequential compositions without additional composition-level evidence and authority conditions. Implementation correspondence is reported separately from the theoretical proposition. Evidence states are explicitly bounded, and undeclared or incomplete implementation evidence is not promoted to VERIFIED status. This release is an external-review preprint (PAPER5-EXTERNAL-REVIEW-FREEZE-004). It does not claim production readiness, universal safety, real-plant validation, independent replication, general runtime effectiveness, or production authorization. The release is intended to provide a stable public reference for external review, discussion, citation, and subsequent evidence-binding revisions.