Governance Laundering: A Taxonomy of Failure Modes in AI Compliance Architectures

AI governance tools are often purchased as if they provide authorization-layer assurance, yet many produce governance artifacts (logs, dashboards, attestations, certificates) that cannot support independent verification of a specific AI decision. This paper names the resulting condition governance laundering: a governance claim whose assurance exceeds the evidence or enforcement properties supporting it. Policy theater and six related failure-mode families describe its structural forms. The paper defines seven evidence-grade governance requirements (R1 through R7), classifies each failure-mode family by the requirement it violates, and formalizes the Expanded Anti-Laundering Protocol (EALP): seven diagnostic tests, extending the Anti-Laundering Tests of A Taxonomy of AI Governance Approaches: Distinguishing Visibility, Alignment, and Authorization (Zenodo DOI: 10.5281/zenodo.18275969), with explicit procedures, pass criteria, and graded outcomes. The paper contributes: A definitional boundary between governance artifacts (bare activity records) and governance evidence sufficient for independent verdict reconstruction A seven-family structural taxonomy of failure modes, from policy theater to semantic drift The EALP: seven tests covering export, offline replay, state completeness, fail-closed enforcement, mutation and drift, pre-execution dependency, and custody of the authorized action A three-grade outcome model (Evidence-Grade, Monitoring-Grade, Artifact-Only), with a withheld-grade rule for incomplete evaluations and stated limits on what each grade establishes Supplemental assessments, recorded separately from the grade, for path coverage (non-bypassability), authority and delegation, and input provenance An a priori susceptibility map across eight categories of governance products, with implications for procurement, audit, and regulatory evaluation Scope. The framework is vendor-neutral and architecture-focused. It evaluates claims of per-decision authorization assurance; it does not propose a governance architecture or treat organizational controls as deficient for serving a different function. Susceptibility, an observed test failure, and a governance-laundering finding are distinct: a laundering finding requires an identifiable representation whose claimed assurance materially exceeds what was established for the same scope. Companion file. This record includes the EALP Evaluation Worksheet, a procurement-ready form for recording the declared scope, the seven tests, the supplemental assessments, the grade, and any representations assessed or evidence outstanding. Update (July 2026, v1.1): Closed the EALP grading rule so that every test outcome maps to a grade. Aligned citations to the AI Governance Taxonomy v1.7 and the Five Tests Standard (5TS) v1.2.0, and added a Provenance deferral note: 5TS adds Provenance (established origin of the inputs grounding a verdict; origin, not truth) as a fifth normative test; the corresponding failure mode and its EALP diagnostic are deferred until input-origin binding is specified, and are identified as the eighth candidate property. Added a terminological note distinguishing governance artifacts (bare activity records) from authorization artifacts (conforming evidence objects). The failure-mode taxonomy, the seven EALP tests, and the susceptibility analysis are unchanged. Update (July 2026, v1.1.1): Consolidates two parallel v1.1 working revisions into a single release. Renamed requirement R6 to Continuity, with content-hash binding as the base requirement and cryptographic-grade binding where external or adversarial verification requires it. Revised the Related Work characterization of adjacent agentic frameworks and compressed the related-instruments discussion connecting the EALP to the authorization artifact gap. Replaced the withdrawn SEC Predictive Data Analytics proposal with the current banking model-risk guidance (SR 26-2 / OCC Bulletin 2026-13 / FIL-15-2026) and added statute citations for New York’s RAISE Act and California’s SB 53 (TFAIA). Added the CC BY 4.0 license block, suggested citation, and archival metadata. The failure-mode taxonomy, the seven EALP tests, and the susceptibility analysis are unchanged. Update (September 2026, v1.2): Methodological clarification and doctrine-alignment edition. The seven failure-mode families, seven test IDs, three grade labels, and susceptibility matrix are retained; procedures, pass criteria, and assurance limits are revised. Evidence-Grade is delimited to the seven tests within a declared scope. Adds evaluator-failure and authorized-resolution checks (T-4), different-version and same-identifier mutation detection (T-5), authorization-dependency testing (T-6), and action-context binding (T-7). Path coverage, authority, and provenance are assessed separately; susceptibility, test failure, and laundering findings are distinguished; claims not substantiated during an evaluation are recorded without being treated as test failures. References are updated to current editions. Correction: earlier versions stated that the EALP Evaluation Worksheet had been released as a companion file; it was not included in the v1.1 or v1.1.1 records and is included from this version. Part of the FERZ research corpus: FERZ community on Zenodo.