A three-stream, auditable screening architecture can plausibly triage customs‑observable trade‑based money‑laundering risks without claiming proven performance; its effectiveness hinges on legal access to gateway evidence and on sentinel-enabled, score‑independent evaluation to overcome selective‑labels bias.

Main Finding

The paper proposes a plausibility-first, institutionally tiered screening architecture for the customs-observable subset of trade-based money laundering (TBML). It extends Kupatadze’s macro–meso–micro framing into three parallel, non‑exclusionary evidence streams (macro corridor‑commodity, meso declaration anomalies, heterogeneous entity‑graph) that combine via late, decomposable fusion. The design is disciplined by an explicit typology mapping TBML techniques to observability, a candidate operator set for each stream, a legal/tier vocabulary (public, customs‑internal, gateway‑conditional) instantiated for Bangladesh, and a three‑stage evaluation pathway built to confront the selective‑labels problem. The paper claims screening plausibility, not demonstrated effectiveness, and specifies how effectiveness would need to be established.

Key Points

• Architecture
  • Three parallel streams run across the full universe of import/export declarations:
    • Macro: corridor × commodity patterns (raises ordering/priority but does not exclude cases).
    • Meso: declaration‑level anomaly detectors.
    • Micro/entity graph: heterogeneous graph stream capturing relationships and sequence patterns.
  • Late, decomposable fusion combines stream outputs so scores remain auditable and components interpretable.
  • A stratified sentinel stream preserves observations that are independent of screening scores to address biased label collection.
• Detectability & Scope
  • The paper provides a typology mapping ten TBML techniques to:
    • Which observables (customs, registry, financial gateway) they leave,
    • Which layer(s) (macro/meso/micro) can detect them,
    • Where the blind spots lie.
  • This makes explicit which TBML modes are amenable to customs‑observable screening and which are effectively invisible.
• Institutional & Legal Tiers
  • Introduces a tier vocabulary separating evidence into public‑data, customs‑internal, and gateway‑conditional layers.
  • Presents a statute‑level legal mapping for Bangladesh to illustrate how evidence access and admissibility constraints shape feasible operators and pipelines.
• Evaluation Pathway
  • Three‑stage prospective evaluation: operational‑baseline analysis → shadow mode → authorised sentinel pilot.
  • Designed to confront the selective‑labels/enforcement bias problem by maintaining score‑independent sentinel observations and a staged permissioning approach.
• Empirical anchoring and examples
  • Uses published institutional evidence (FIU reporting concentration, a development‑bank pilot on trade‑related suspicious reporting, and external gap estimates) as contextual anchoring for screening design—explicitly not as laundering prevalence measures.
  • Provides three worked illustrations spanning under‑invoicing (revenue loss) to outward value transfer.
• Claims & limits
  • The framework executes no predictive model; it demonstrates architectural plausibility and prescribes how performance claims must be earned through the specified evaluation path.
  • Explicit about blind spots and displacement risks (situational‑prevention/crime‑script reading).

Data & Methods

• Nature of the work: conceptual / systems design paper (no empirical model fit; no deployed classifier).
• Components and methods used:
  • Typology‑detectability scope mapping: catalogues ten TBML techniques and maps them to observables, layers, and blind spots.
  • Operator specification: proposes candidate operators/algorithms for each stream (e.g., aggregate corridor statistics, declaration anomaly scores, graph‑based relationship detectors), with emphasis on decomposability and auditability.
  • Fusion design: late fusion of stream outputs into decomposable scores, preserving provenance and interpretability of sub‑scores.
  • Sentinel & evaluation design: operational baseline, shadow testing, and authorised sentinel pilot to produce (selectively) unbiased labels and measure performance while respecting legal constraints.
  • Institutional/legal analysis: constructs a three‑tier evidence vocabulary and applies a statute‑level mapping for Bangladesh to illustrate constraints on data flows and gateway evidence.
  • Worked examples: three illustrative case walks showing how the architecture would surface directional signals across the spectrum of TBML techniques.
• Data referenced (as contextual anchors, not measurements):
  • FIU reporting concentration statistics.
  • Development‑bank pilot results on trade‑reporting.
  • External estimates of trade reporting gaps.

Implications for AI Economics

• For model design and deployment
  • Late, decomposable fusion and per‑stream provenance are important for auditability, regulatory explainability, and for constructing evaluation plans robust to selective labels.
  • Stratified sentinel design is a practical operational requirement to obtain score‑independent labels and to avoid feedback loops that bias training/evaluation data.
  • Legal/tier constraints (public vs customs‑internal vs gateway‑conditional evidence) must be encoded into data access, feature engineering, and downstream cost models — economics of inference must include evidence acquisition costs and admissibility risk.
• For measurement of illicit flows and economic policy
  • The typology clarifies which TBML techniques are attention‑worthy for customs data analysis and which require complementary interventions (financial gateways, registries, international cooperation).
  • Because the framework is explicit about blind spots, policy evaluation should incorporate displacement risk and the marginal cost of closing specific detection gaps versus expected revenue recovery or crime‑prevention benefits.
• For empirical research & evaluation
  • Researchers should prioritize building realistic evaluation ecosystems (shadow mode + sentinel) rather than relying on enforcement labels alone; selective‑labels correction is critical for unbiased performance assessment.
  • Opportunity for simulation‑based benchmarking: simulate heterogeneous TBML behaviours under varying access tiers to quantify tradeoffs and expected utility of different screening architectures.
  • Causal and cost‑benefit analyses are needed to move from plausibility to effectiveness: estimate how screening raises offender effort/risk, enforcement resource allocation, and displacement patterns.
• For broader AI economics questions
  • Illustrates how institutional constraints shape algorithmic design choices and the economics of information: access rights, legal admissibility, and interagency cooperation alter the feasible feature set and thus detection frontier.
  • Highlights the importance of decomposability for regulatory transparency and for aligning incentives across institutions (customs, FIU, banks).
  • Points to open research on optimal allocation of inspection/enforcement budgets under detection uncertainty and partial observability.

Limitations to keep in mind - Conceptual only: no empirical validation or claims of operational effectiveness. - Focused on customs‑observable TBML; many techniques remain outside scope or require gateway cooperation. - Legal mapping illustrated for Bangladesh — transferability requires redoing the statute/tier mapping for other jurisdictions.