Compliant AI Infrastructure for Regulated Finance: A tiered multi-agent framework with DLT audit trails for financial operations in DACH

Summary

This paper proposes a reference architecture for deploying AI in regulated financial institutions across the DACH region (Germany, Austria, Switzerland), arguing that compliance should be embedded at execution time rather than audited ex post. Its core device is a Regulatory Intent and Exposure Matrix: every legal provision is reduced to a label v = (t, e), with intent being Prohibition or Obligation and exposure Internal or External. A governed policy compiler maps labels and task context into selected prohibitions, obligations, an evaluation order, dominant disqualifiers and runtime budgets. Prohibitions act as feasibility constraints that can block output externalisation; obligations extend tasks with evidence artefacts that must pass an admissibility predicate covering provenance, signatures, trusted time, DAG inclusion, retention and access. Committee activation is policy-driven and proportionate, collapsing to a compliance gate when dominant disqualifiers fire. All decisions, reason codes and evidence hashes are anchored to a permissioned DAG ledger supporting replay and auditor verification. The formalism spans Equations (1)-(17) plus two pages of stylised pseudocode, illustrated by a cross-border retail order example. The authors state the contribution is deliberately theoretical and algorithmic: there is no implementation, dataset, experiment or quantitative evaluation, determinism proofs are deferred, and the legal-to-matrix classifier is unvalidated. Key limitations are the absent empirical evidence, a binary applicability predicate that may become a single point of failure, and unverified portability claims.