PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Walter Kurz, Reinhard Magg

arXiv:2609.27632v2q-fin.GNcs.AI

TL;DR

Financial institutions lack architectures that embed supervisory requirements for accountability, resilience, and auditability into AI execution. The paper proposes a policy-compiled, compliance-first architecture that treats regulation as orientation rather than a deterministic ruleset. Its theoretical design binds admissible execution evidence to a permissioned DAG, while its authors identify large-scale empirical evaluation as future work.

  • Problem

    The study addresses the lack of AI architectures that integrate supervisory expectations for accountability, resilience, and auditability into system design.

    Source: §1 · p. 1

  • Method

    A regulatory intent-and-exposure matrix guides a policy compiler that selects prohibitions and obligations, while admissibility checks and a permissioned DAG govern and record execution.

    Source: §4.1 · p. 4, §4.2 · p. 7, §5 · p. 12

  • Results

    The paper presents a theoretical architecture in which prohibitions constrain feasible execution, obligations require admissible evidence, and DAG records support replayability, provenance checks, and failure attribution.

    Source: §4.2 · p. 8, §6 · p. 13

  • Takeaways & Limitations

    Separating portable regulatory orientation from institution-specific activation offers a design pattern for compliance assurance across DACH and the wider EU.

    Source: §5 · p. 12, §6 · p. 13

    The contribution is theoretical and algorithmic; meaningful empirical evaluation requires full system deployment across institutional and regulatory contexts.

    Source: §7 · p. 13

Abstract

from arXiv · show

We present a compliance-first architecture for AI in regulated finance that treats regulation as an orientation layer rather than a deterministic ruleset. A matrix of regulatory intent and exposure provides a compact classification handle, which a governed policy compiler then maps into concrete prohibitions, obligations and runtime budgets. Prohibitions constrain feasibility and block externalisation, while obligations extend tasks with artefacts that must meet explicit admissibility criteria. Committee activation remains policy-driven and proportionate, preserving efficiency while ensuring supervisory oversight. Evidence, decisions and reason codes are bound to a permissioned DAG with deterministic timestamping, enabling replay, provenance checks and clear attribution of failure. Clause-level legal indexing with effective dates and capability-based agent routing ensure portability across DACH and the wider EU. The result is assurance by construction: compliance is embedded in execution and verifiable by auditors without sacrificing proportionality or transparency.

1 Introduction

The study proposes a reference architecture for compliant AI in finance, addressing the gap between AI adoption and supervisory expectations for accountability, resilience, and auditability. It focuses on DACH institutions, whose regulatory convergence and institutional diversity provide a testing ground.

1 Introduction

  • The study targets the lack of architectures that integrate supervisory expectations for accountability, resilience, and auditability into AI system design.It focuses on financial institutions in the DACH region.Source: §1 · p. 1
  • DACH offers a coherent yet diverse testing ground, with convergent prudential expectations and reporting cultures alongside varied institutional types.The region's framework supports methodological consistency and comparison across jurisdictions.Source: §1 · pp. 1, 3
  • The study defines financial institutions as supervised entities across deposit-taking, credit, investment, insurance, asset-management, payment, settlement, and market-infrastructure activities.The introduction situates this scope within a broad EU and DACH regulatory corpus.Source: §1 · p. 2

2 Related Work

Prior research addresses regulated finance, financial multi-agent systems, audit logging, and distributed ledgers, but the paper identifies a gap in integrating these strands into a supervision-ready AI architecture.

2 Related Work

  • Financial multi-agent research has advanced in trading, portfolio optimisation, and credit risk, but typically assumes static policies and lacks verifiable execution evidence.Examples include hierarchical deep reinforcement learning for portfolio optimisation Sun, 2025 and multi-agent dynamic portfolio learning Cheng, 2024.Source: §2 · p. 3
  • Audit and model-risk research stresses traceability and oversight, while cryptographic timestamping remains uncommon as a first-class principle in multi-agent orchestration.DAG-based audit-grade logging for AI coordination is described as under-explored.Source: §2 · p. 3
  • The paper identifies no existing enterprise architecture that combines compile-time and run-time compliance constraints, gated multi-agent coordination, and deterministic supervision-ready evidence.The gap is especially salient for tiered designs serving institutional and consumer interfaces under DACH norms.Source: §2 · p. 3

3 Contribution

The contribution is a DACH-oriented reference architecture that embeds machine-checkable regulatory duties in AI orchestration and execution. It combines tiered, role-scoped agents with gated actions and permissioned-ledger evidence.

3 Contribution

  • The architecture encodes regulatory duties as machine-checkable constraints enforced before and during execution through a policy store and rule engine.It is aligned with DACH supervisory context and designed for transfer across the EU.Source: §3 · p. 3
  • Agent roles are scoped across institutional and consumer tiers, with explicit permissions, explainability thresholds, and segregation of duties.Sensitive actions such as order placement require dual control, and each step produces audit-ready evidence.Source: §3 · p. 4
  • A permissioned DAG provides deterministic timestamping and inclusion proofs for routing instructions, budget usage, and control attestations.The ledger supports verification without imposing settlement semantics on application workflows.Source: §3 · p. 4

4 System Architecture

Section 4 describes an architecture that embeds compliance in execution by compiling regulatory orientation and task context into governed policy checks, with proportionate oversight and verifiable audit events recorded on a permissioned DAG. It defines admissibility requirements for obligation evidence and presents replayable compilation in which checks may lead to denial, alongside clause-level legal indexing for portable, reviewable operation across jurisdictions.

4 System Architecture

  • A permissioned DAG ledger records agent actions, constraint evaluations, and gating decisions with deterministic timestamps and inclusion proofs for audit and replay.The ledger is a compliance substrate for evidence packs, routing instructions, and oversight signatures, not a mechanism for financial transaction settlement.Source: §4 · p. 4

4.1 Agnostic Regulatory Intent and Exposure Matrix

  • The Regulatory Intent and Exposure Matrix labels provisions as prohibitions or obligations and internal or external, orienting reasoning while policy compilation selects concrete gates.Prohibitions can block externalisation; obligations can require admissible evidence, while human oversight may revise gate selection.Source: §4.1 · pp. 4, 6

4.2 Prohibitions as feasibility and obligations as task extension

  • Policy compilation makes prohibitions feasibility constraints and obligations task extensions requiring admissible evidence, with short-circuit evaluation and DAG records enabling auditable gate decisions.The audit record supports replay and independent re-evaluation of selected checks, evidence admissibility, and the resulting decision.Source: §4.2 · pp. 7–8

4.3 Agent activation by intent and exposure

  • Compiled policy and oversight activate a single agent or minimal committee according to role needs, exposure, selected checks, and runtime budgets, limiting unnecessary coordination.The orchestration handshake returns artefacts and rationales for a compliance agent’s pre-output gate; committee membership and decisions are recorded to the DAG.Source: §4.3 · pp. 8–9

4.4 Legal corpus indexing and agent routing

  • Clause-level legal indexing with canonical citations and effective dates, combined with capability-based agent routing, supports jurisdiction-aware retrieval and portable supervisory review.Retrieval coordinates semantic, citation, and graph-based views, while the design avoids statute-specific agent silos and reliance on embeddings alone to correlate text with law.Source: §4.4 · pp. 9–10

4.5 Evidence admissibility model

  • Obligation evidence is audit-ready by construction: artefacts carry provenance, signatures, trusted time and DAG inclusion, while raw records remain in controlled storage.Admissibility also requires policy-compliant retention and role-scoped access with appropriate redaction.Source: §4.5 · p. 10

4.6 Policy compilation and governance

  • The matrix narrows candidates but does not activate checks; a versioned compiler selects ordered guards, runtime budgets and admissibility-bound obligation templates using context, jurisdiction, exposure and institutional policy.Budget breaches may trigger committee escalation or deny externalisation.Source: §4.6 · pp. 10–11
  • Signed compilation certificates anchored to the DAG support replay, while controlled overrides require reason codes and dual control and cannot weaken dominant prohibitions for external exposure.The compiler is deterministic for given orientation, context, policy version and time; output changes can therefore be attributed to changes in those inputs or time-scoped rules.Source: §4.6 · p. 11

4.7 Stylised simulation pseudocode

  • The pseudocode illustrates context detection, policy compilation, disqualifier and admissibility checks, and externalisation gating, including denials for budget breaches or dominant prohibitions.The supplied excerpt shows the two denial records but only a partial continuation of the pseudocode.Source: §4.7 · p. 11

20: if χ(τ)̸ = 0 then

The pseudocode detects applicability, compiles policy, evaluates feasibility and obligation admissibility, gates externalisation, and records decision paths in a replayable audit pack.

20: if χ(τ)̸ = 0 then

  • When obligation evidence is inadmissible, the procedure records a denial with the failed obligation as its reason.Source: §20 · p. 12
  • The procedure forms the feasible policy set by requiring prohibitions to be false and obligation evidence to be admissible.Source: §20 · p. 12
  • If no policy is feasible, the procedure records a denial for lack of a feasible policy.Source: §20 · p. 12
  • The procedure records every decision path in a replayable, time-bound audit pack after applying the externalisation gate.The pack supports later reconstruction of the decision path.Source: §20 · p. 12

5 Discussion

The discussion frames the architecture as separating regulatory orientation from institution-specific policy activation. It emphasizes auditability and portability, while identifying curation, privacy, and determinism constraints.

5 Discussion

  • Regulatory constraints are first-class objects, but the intent–exposure matrix orients compilation rather than determining execution.Source: §5 · p. 12
  • Prohibitions constrain feasibility, while obligations require admissible evidence and do not trade against utility.Obligation artefacts must satisfy the admissibility predicate E(·).Source: §5 · p. 12
  • Policy-driven committee activation and short-circuiting of dominant checks support proportionality and reduce latency without sacrificing audit-trail completeness.Source: §5 · p. 12
  • Misclassification or overly broad dominance sets may cause false blocks, while incomplete policy stores may cause false releases.The legal index and intent–exposure mapping require continuing curation, versioning, and reviewer sign-off.Source: §5 · p. 12
  • DAG privacy constraints require anchoring hashes and metadata while keeping raw artefacts in controlled storage with disciplined key management.Source: §5 · p. 12
  • Determinism depends on stable time sources and reproducible compilation; the design does not replace legal interpretation or product governance.Portability across DACH and the wider EU follows from separating a portable orientation layer from institution-specific activation.Source: §5 · p. 12

6 Conclusion

The conclusion presents a compliance-first architecture that uses regulatory intent and exposure to orient classification and governed policy compilation. It combines runtime constraints and proportionate oversight with a timestamped audit trail and jurisdictional portability.

6 Conclusion

  • The Regulatory Intent and Exposure Matrix classifies and routes tasks, while a governed policy compiler maps regulation into concrete controls.The architecture treats regulation as an orientation layer rather than a deterministic ruleset.Source: §6 · p. 12
  • Prohibitions constrain feasibility and block externalisation, while obligations add artefacts that must satisfy admissibility criteria.Optimisation proceeds over feasible options under lexicographic semantics.Source: §6 · pp. 12–13
  • Policy-driven committee activation supports proportionality, and dominance rules allow early termination without compromising audit completeness.Source: §6 · p. 13
  • A permissioned DAG with deterministic timestamping binds evidence, decisions, and reason codes for replay, provenance checks, and failure attribution.Portability across DACH and the wider EU is supported by orientation–activation separation, a clause-level legal corpus, and capability-based routing.Source: §6 · p. 13

7 Limitations and Future Research

The design depends on curated legal mappings, reliable policy compilation, and institution-specific calibration of admissibility and governance choices. Its empirical evaluation remains theoretical and algorithmic pending full-scale deployment across institutional and regulatory contexts.

7 Limitations and Future Research

  • Curated legal mappings and policy compilation can drift or contain gaps, causing false blocks or unintended releases.Source: §7 · p. 13
  • Compiler determinism and reproducibility require explicit proof obligations and differential testing.Source: §7 · p. 13
  • Admissibility criteria need institutional calibration, while ledger privacy, cryptographic choices, and key management affect audit reliability.Source: §7 · p. 13
  • Committee thresholds, runtime budgets, legal retrieval robustness, and reviewer workflows need empirical calibration, testing, and ongoing governance effort.Future work includes institution-specific DACH evaluation, auditor-facing query suites, and privacy-preserving ledger and time-binding designs.Source: §7 · p. 13
  • Meaningful empirical evaluation requires full system deployment across institutional and regulatory contexts, beyond the scope of this academic research.Smaller prototypes may misrepresent systemic properties of budget enforcement, admissibility, or externalisation.Source: §7 · p. 13

Declarations

The declarations report authorship, funding, conflicts of interest, data availability, and editorial use of generative AI.

Declarations

  • All authors contributed equally; the research received no external funding, and the authors declared no conflicts of interest.Source: Declarations · p. 16
  • Supporting data and code are available from the corresponding author on reasonable request.Source: Declarations · p. 16
  • Generative AI tools were used for editorial work, and the authors verified the content and retain responsibility.Source: Declarations · p. 16
Loading 2609.27632v2…