Detecting Trade-Based Money Laundering in the Documents Themselves: A Cross-Validation Graph with Closed-Loop Self-Iteration

Shijie Qu

KEOI Research Institute

sj@keoi.org

KEOI Working Paper · October 2026

Abstract

Trade-based money laundering (TBML) hides in the documents of trade rather than in the payments that settle it. Conventional transaction-monitoring systems examine amounts, frequencies, and velocities, and are therefore blind to the documentary substance of a shipment; machine-learning systems trained on transaction-level features inherit the same blindness; and static models of either kind decay as laundering typologies migrate across goods, corridors, and forgery techniques. This paper describes a detection system, disclosed in U.S. Provisional Patent Application No. 64/170,275, that is organised around the documents themselves. Heterogeneous trade documents — commercial invoices, bills of lading, customs declarations, packing lists, letters of credit, and associated payment records — are parsed into canonical shipment assertions that retain provenance metadata, and entity mentions across documents are resolved to canonical identities, with alias collisions preserved as signals rather than merged away. The resolved material is assembled into a heterogeneous entity–document–shipment graph whose edges are of two kinds: documentary edges recording what each document declares, and consistency edges — computationally verifiable assertions linking attributes across documents that purport to describe the same shipment — each carrying a consistency score. From this graph the system computes features that encode known TBML typologies at the documentary level: unit-price deviation against benchmark price distributions indexed by HS code, trade corridor, and time window; cross-document quantity and value reconciliation; circular-trading cycles; phantom-shipment document asymmetries; and shipment-structuring patterns. A calibrated scoring model emits per-shipment and per-entity risk scores with machine-readable documentary reason codes, exposed through a real-time scoring API designed for pre-settlement interception. Investigation outcomes — suspicious activity report filings, confirmed typologies, and analyst labels — close the loop: confirmed activity reweights the suspicious-entity subgraph and re-runs community detection, labelled cases retrain the model, an active-learning strategy allocates analyst review by uncertainty over graph neighbourhoods, and per-corridor stationarity testing of benchmark price distributions detects typology drift and triggers benchmark re-estimation, targeted labelling, and model refresh — all without manual feature re-engineering. A further embodiment extends the graph across the full lifecycle of direct-to-consumer cross-border e-commerce. The paper reports no evaluation and no measurements; numerical values cited are the design targets and structural parameters stated in the disclosure.

Keywords: trade-based money laundering; trade document cross-validation; entity resolution; heterogeneous graph; closed-loop learning; active learning; concept drift; financial crime detection

1. Introduction

Every anti-money-laundering system makes an implicit choice about where laundering is visible. The conventional choice is the payment: monitoring systems observe transactions as they move through accounts and look for amounts, frequencies, and velocities that depart from expected patterns. For trade-based money laundering (TBML) — the disguising of illicit funds as legitimate trade flows through mis-invoicing, phantom shipments, and circular trading, recognised by the Financial Action Task Force (FATF) as one of the principal channels for moving criminal proceeds across borders (FATF, 2006) — that choice is badly matched to the crime. The payment that settles a laundered shipment is, in itself, usually unremarkable: a commercial transfer between two firms, of a plausible size, at a plausible frequency. What is false about the transaction is not its shape in payment space but its relationship to the goods it purports to pay for — and that relationship is recorded not in the payment but in the trade documents that surround it: the commercial invoice, the bill of lading, the customs declaration, the packing list, the letter of credit. Each of these documents describes the same underlying shipment from an independent institutional vantage point — seller, carrier, customs broker, bank — and it is their mutual inconsistency, not any single document's content, that constitutes the evidentiary core of TBML.

This paper describes a detection system, disclosed in U.S. Provisional Patent Application No. 64/170,275 (filed October 6, 2026), that takes the documents as its primary object. The system fuses heterogeneous trade documents into a single machine-analysable structure — an entity–document–shipment graph in which cross-document consistency assertions are computationally verifiable edges with scores — engineers features from that structure that encode TBML typologies at the documentary level, and closes a learning loop in which investigation outcomes continuously refine the detection models at the level of document semantics. The disclosure frames the technical problem as threefold: how to fuse heterogeneous trade documents into one structure that exposes cross-document inconsistencies; how to engineer TBML-specific features from that structure that go beyond transaction-level statistics; and how to close the learning loop so that outcomes — suspicious activity reports (SARs), confirmed typologies, analyst labels — refine the models without manual re-engineering. The paper follows that organisation. Section 2 sets out why payment-centric monitoring fails on TBML and what the frontline institutions require instead. Sections 3–8 describe the system: ingestion into canonical shipment assertions, entity resolution, the cross-validation graph, the typology feature families, and calibrated scoring with documentary reason codes. Section 9 describes the closed-loop self-iteration machinery — feedback propagation, retraining, active learning, and per-corridor drift detection. Sections 10–12 cover the real-time scoring API, the full-lifecycle e-commerce embodiment, and alternative embodiments. Section 13 discusses the claimed technical effects and the paper's limitations, and Section 14 concludes. Throughout, the paper confines itself to the system as disclosed in the patent filing; it reports no experiments, no evaluation, and no deployment results.

2. Background: A Crime Conducted in Documents

TBML is not a niche technique. Public estimates of trade mis-invoicing involving developing economies run into the hundreds of billions of dollars annually, and the proceeds moved through trade channels derive from underlying crimes — narcotics trafficking, human trafficking, corruption, terrorist financing — whose social cost gives detection its urgency (FATF, 2006; FATF & Egmont Group, 2020). The frontline responsibility falls on institutions that each see part of the documentary record: banks and payment institutions, which settle the underlying payments; trade-finance providers, which handle letters of credit and documentary collections; customs authorities, which clear the goods; and insurers, which underwrite cargo and trade-credit risk.

Against this problem, existing technology fails in three distinct ways, and the disclosure is explicit about each.

The first failure is payment-centrism. Conventional AML transaction monitoring was designed for payment flows. Its features — amount, frequency, velocity — are blind to documentary substance. A launderer who grossly over-invoices a shipment produces a payment that is, in amount–frequency–velocity space, an ordinary commercial transfer. Rule-based attempts to catch such cases from transaction statistics generate overwhelming false positives, because legitimate trade routinely exhibits wide price dispersion across corridors, seasons, and counterparties: price variation is what trade looks like, and a statistic that cannot separate dispersion from markup cannot separate commerce from laundering.

The second failure is feature-level blindness in generic machine learning. Models trained on transaction-level features cannot exploit the richest available signal, the trade documents themselves. The documents describe one shipment from several independent vantage points, and their inconsistencies — quantity against quantity, declared value against declared value, goods described one way on a packing list and another way on a declaration — are precisely where mis-invoicing, phantom shipments, and circular trading leave traces. No conventional system constructs a unified cross-validation structure over these documents; the inconsistencies remain, at best, manual review notes assembled case by case by an analyst who must first suspect the shipment for other reasons.

The third failure is stasis. Laundering typologies evolve: as one mis-invoicing corridor is understood and watched, value moves to new goods, new routes, new document-forgery techniques. Static models — even accurate ones — decay. Existing self-learning AML addresses staleness only at the level of generic transaction features, re-weighting risk factors after case closures or detecting statistical drift in transaction streams. None anchors its learning loop in trade-document semantics, which is where TBML actually mutates: the abused HS codes change, the corridor benchmark distributions shift, the forgery techniques alter which cross-document assertions fail. A detection system for TBML must therefore learn at the level at which the crime is conducted — the level of documents — or accept that its knowledge expires.

These three failures dictate the design requirements the disclosed system answers: a single structure that fuses the documents and exposes their inconsistencies as computable objects (Sections 4–6); features engineered from that structure that encode typologies documentarily rather than transactionally (Section 7); and a learning loop closed over investigation outcomes at the document-semantic level (Section 9).

3. System Overview

The disclosed system (FIG. 1 of the filing) is organised around three interlocking technical contributions.

The first is the multi-source trade-document cross-validation graph. An ingestion layer normalises heterogeneous trade documents — commercial invoices, bills of lading, customs declarations, packing lists, letters of credit, and associated payment records — into a unified entity–document–shipment graph. Nodes represent entities (exporters, importers, carriers, banks, customs brokers), documents, shipments, and payments. Edges are of two kinds: documentary edges, which encode what a document declares (an invoice declares goods at a unit price; a bill of lading attests a shipment of a quantity), and consistency edges, which encode cross-document assertions that are computationally verifiable — that the quantity on the invoice equals the quantity on the bill of lading, that the invoice's unit price lies within the corridor benchmark distribution, that the customs-declared value approximates the invoice value. Inconsistencies — quantity mismatches, price deviations, entity-alias collisions — become first-class graph features rather than manual review notes.

The second contribution is TBML-typology feature engineering anchored on document semantics. From the graph, the system computes features that directly encode known laundering typologies at the documentary level: unit-price deviation against per-HS-code, per-corridor benchmark price distributions; cross-document quantity and value reconciliation scores; circular-trading indicators from directed cycles in the entity–shipment subgraph; phantom-shipment scores from document-existence asymmetries; and structuring indicators from shipment-splitting patterns. A risk-scoring model consumes these features to produce per-shipment and per-entity risk scores with machine-readable reason codes stated in documentary terms.

The third contribution is closed-loop self-iteration at the document-semantic level. Investigation outcomes are ingested as labelled feedback and propagated through the graph: confirmed illicit activity updates the edge weights and community structure of the suspicious-entity subgraph, and labelled cases retrain the scoring model on document features (FIG. 2 of the filing). An active-learning query strategy selects the cases whose review would be most informative, and a per-corridor concept-drift detector monitors the stationarity of benchmark price distributions, triggering model refresh when techniques mutate. The loop requires no manual feature re-engineering: because the features are anchored on document semantics, retraining re-calibrates the documentary signatures themselves — which HS codes are currently abused, which corridors' benchmarks have shifted — as typologies evolve.

A real-time scoring API exposes the score and reason codes to subscribing institutions — banks, payment processors, trade-finance platforms — enabling pre-settlement interception of high-risk trade payments. Because the scores arrive accompanied by documentary reason codes, the output is designed to satisfy the explainability expectations of financial-crime compliance functions: a score that cannot say, in documentary language, why it fired is of limited use to the analyst who must act on it.

4. Document Ingestion and Shipment Assertions

The ingestion layer (110) receives trade documents from multiple institutional sources: commercial invoices, bills of lading, customs declarations, packing lists, letters of credit, and associated payment records, including SWIFT messages and correspondent-banking records in the disclosure's enumeration. Each document is parsed into a canonical schema of shipment assertions — structured statements recording the declaring party, goods description, HS code, quantity, unit price, currency, counterparties, dates, and transport identifiers. The assertion, not the document image or text, is the unit on which everything downstream operates: it is the form in which a document's claims about a shipment become comparable with every other document's claims about the same shipment.

Parsing is matched to document form. Structured forms, such as customs declarations, are handled by template-based extraction; scanned documents by optical character recognition with layout analysis; free-text descriptions by natural-language field extraction. Every assertion retains provenance metadata — source document, source institution, ingestion timestamp. Provenance is not bookkeeping decoration. A consistency assertion between two documents is only meaningful if the documents are genuinely independent attestations; an invoice and a payment record generated from the same underlying submission do not corroborate each other the way a carrier's bill of lading corroborates a seller's invoice. Retaining the source of every assertion is what allows the graph, later, to treat agreement and disagreement as evidence rather than as noise.

5. Entity Resolution and the Alias Signal

Before documents can be cross-validated, the parties they mention must be identified across documents (120). The same exporter appears on an invoice under one name form, on a bill of lading under an abbreviation, and in a payment record under a transliteration. The disclosed resolution is hybrid: deterministic keys — tax identifiers, and Legal Entity Identifier codes where available — anchor identities where they exist, and probabilistic record linkage, name and address fuzzy matching with blocking on country and trade corridor, resolves the remainder.

The design decision that distinguishes this stage is its treatment of failure. Alias collisions — the same legal entity appearing under variant names across documents — are flagged as first-class signals rather than silently merged. The reasoning is typological: deliberate aliasing is itself a TBML indicator, since a laundering structure benefits from fragmenting its own documentary footprint, while legitimate name variation is usually stable and explicable. A resolution process that quietly normalises every variant into one identity destroys exactly the evidence that distinguishes the two cases. The disclosed system therefore resolves entities for the purpose of graph construction while preserving the collision as a feature: the graph knows that these names are one entity, and it also knows that the entity presents itself under several.

6. The Cross-Validation Graph

The graph builder (130) assembles resolved entities, documents, shipments, and payments into a heterogeneous graph. Its edges are of the two kinds introduced in Section 3. Documentary edges record declared relations: an invoice declares a shipment, a bill of lading attests it, a payment settles an invoice. Consistency edges (FIG. 3 of the filing, which illustrates the graph for a single shipment) are computationally verifiable assertions linking attributes across documents that purport to describe the same shipment: invoice quantity equals bill-of-lading quantity; invoice unit price lies within the corridor benchmark distribution; customs-declared value approximates invoice value.

Each consistency edge carries a consistency score in [0, 1], computed by a comparator specific to the assertion type: exact match for identifiers; tolerance-banded numeric comparison for quantities; a distributional test for prices. The score is the paper's central data object, and its design deserves emphasis. A binary pass/fail assertion would force every tolerance question to be settled in advance, in the comparator, where it cannot be revisited; a score defers the judgement to the feature and model layers, where the significance of a small quantity discrepancy can be weighed against the rest of the shipment's evidence. A distributional comparator likewise refuses the fiction that a price is simply right or wrong: it asks how improbable the declared price is, given what the corridor ordinarily charges for the goods — which is the form in which the over-invoicing question actually arises.

The graph is also the unit of accumulation. A single shipment's graph (FIG. 3) shows that shipment's documents checking each other; the corpus graph, built shipment by shipment, is what makes entity-level and corridor-level structure visible — the cycles, communities, and benchmark distributions of the following sections are properties of the accumulated graph, not of any one shipment.

7. Typology Features

From the graph, the feature layer (140) computes a feature vector per shipment, and aggregates per entity, across five families that correspond one-to-one to the classical TBML typologies (Sections 7.1–7.5). The correspondence is the design's point: each family exists because a named laundering technique leaves a named documentary trace.

### 7.1 Price deviation

For each invoice line item, the system computes the deviation — a z-score or robust equivalent — of the declared unit price against a benchmark price distribution indexed by HS code × trade corridor × time window, maintained as an exponentially-weighted empirical distribution. Persistent one-sided deviation is the documentary signature of over- and under-invoicing: a single extreme price may be an error or an unusual bargain, but a party whose declared prices sit persistently on one side of its corridor's distribution is exhibiting the pattern by which value is transferred through trade. The benchmark's indexing is doing substantive work here. Prices are only meaningful relative to what the same goods ordinarily cost on the same route in the same period; indexing by HS code, corridor, and time window is what converts raw price dispersion — the false-positive engine of Section 2 — into a deviation measurement.

### 7.2 Cross-document reconciliation

Reconciliation features aggregate the consistency scores across a shipment's consistency edges: quantity mismatches between invoice and bill of lading, value mismatches between invoice and customs declaration, goods-description divergence between packing list and declaration. Where price deviation asks whether the shipment's economics are improbable, reconciliation asks the prior question — whether the shipment, as described by its several documents, is one shipment at all, coherently described. Mis-invoicing at scale is difficult to keep consistent across documents issued by independent parties, and the reconciliation aggregate is where that difficulty surfaces.

### 7.3 Circular trading

Circular-trading features detect directed cycles in the entity–shipment subgraph — A ships to B, B ships to C, C ships back to A — scored by cycle length, margin compression, and temporal density. A cycle with economically implausible round-trip margins, executed densely in time, is the graph form of circular trading: goods (or merely documents) travelling in a loop so that value and apparent trade volume accumulate without economic substance. Cycle detection is a property of the accumulated graph — no single shipment in the loop need look anomalous — which is why this family, like the alias signal, is unavailable to any system that scores transactions one at a time.

### 7.4 Phantom shipments

Phantom-shipment features measure document-existence asymmetries: a payment settling an invoice for which no bill of lading or customs declaration exists in the graph within a plausible time window, or a bill of lading with no corresponding customs entry. The asymmetry is the signal because it is structural: real shipments generate the full documentary set as a by-product of physically moving goods through independent hands, while a fabricated shipment must deliberately manufacture each document, and the seams show in what is missing.

### 7.5 Structuring

Structuring features capture shipment-splitting patterns: multiple sub-threshold shipments between the same entity pair within a short window whose aggregate value crosses reporting thresholds. As with cycles, the pattern exists only in aggregate — each split shipment is individually beneath notice — and the feature is defined at the level of the entity pair over time rather than the individual shipment.

8. Risk Scoring and Documentary Reason Codes

The risk-scoring engine (150) consumes the feature vector and outputs a calibrated TBML risk score per shipment and per entity. In the disclosure's preferred embodiment, the model is a gradient-boosted tree ensemble over the TBML feature vector, with graph-neural-network embeddings of the entity neighbourhood as additional inputs — the neighbourhood embedding carrying into the score the community structure that Section 9's feedback propagation maintains. Calibration is maintained via isotonic regression on held-out labelled outcomes, so that scores are interpretable as probabilities: a score is a statement about likelihood, usable in the threshold and triage decisions compliance functions actually make, rather than an uncalibrated ranking index.

Every score is accompanied by machine-readable reason codes identifying the top contributing features in documentary terms — a price-deviation code naming the HS code and the deviation magnitude in benchmark-distribution units; a quantity-mismatch code naming the disagreeing document pair (the disclosure's example is of the form QTY_MISMATCH_INVOICE_BOL). The reason codes are the interface between the model and the institution. They are designed to be directly consumable by compliance case-management systems, and they are what makes the system's output examinable: an analyst receiving an alert receives, with it, the documentary propositions the alert rests on, each of which can be checked against the documents themselves. In a domain where an unexplained score imposes investigation cost without directing it, the reason code is not an accessory to the score but half of the product.

9. Closed-Loop Self-Iteration

The system's third contribution is the machinery by which it improves (FIG. 2 of the filing). The loop has five stages, numbered 210–260, and its defining property is the level at which it operates: outcomes are fed back not merely as additional training rows but as modifications to the graph and the benchmarks against which future shipments are read.

### 9.1 Outcome ingestion (210)

Investigation outcomes enter as labelled feedback: SAR filings, which supply positive labels with typology tags; analyst accept and reject decisions on reviewed alerts; and confirmed-negative outcomes from cleared investigations. Labels attach at the shipment and entity level, with timestamps. The label set is deliberately three-valued in effect — confirmed illicit, reviewed and rejected, investigated and cleared — because the cleared and rejected cases are what prevent the loop from becoming a self-confirming accumulation of suspicion.

### 9.2 Graph-level feedback propagation (220)

Confirmed illicit activity does not merely become another training row. It updates the suspicious-entity subgraph directly: edge weights between entities co-occurring in confirmed laundering structures are increased, and community detection — Louvain or label-propagation clustering, in the disclosure's examples — is re-run on the updated weights, so that newly confirmed typologies reshape the entity-risk topology itself. This is the mechanism by which the system learns laundering structures and not merely laundering instances: a confirmed circular-trading ring raises the prior risk of its entire graph neighbourhood, including entities whose own documents have not yet attracted attention, because participation in the structure is itself the evidence.

### 9.3 Model retraining (230)

On a scheduled cadence and on drift-triggered events (Section 9.5), the scoring model is retrained on the accumulated labelled corpus, with sample weights emphasising recent confirmed typologies. Because the features are anchored on document semantics, retraining automatically re-calibrates the documentary signatures — which HS codes are currently abused, which document pairs currently disagree in laundering cases — without manual feature re-engineering. The distinction matters operationally: in a conventional pipeline, adapting to a new typology requires an analyst to notice it, a data scientist to engineer a feature for it, and an engineer to deploy the change; here, the feature vocabulary is fixed at the level of document semantics, and what changes with the typologies is the weighting the model places on a vocabulary that already contains the terms in which the new typology is expressed.

### 9.4 Active-learning query strategy (240)

Analyst attention is the loop's scarce resource, and the disclosed strategy spends it by expected information rather than by score order. Cases are selected for review by uncertainty sampling over graph neighbourhoods: shipments whose risk score carries high model uncertainty and whose graph neighbourhood is information-rich — many unresolved consistency edges, proximity to known suspicious communities — are prioritised, so that each analyst label maximally reduces uncertainty over the surrounding subgraph rather than merely resolving the case at hand. In the preferred embodiment, the query budget is allocated per trade corridor in proportion to observed drift magnitude (Section 9.5), concentrating review effort where typologies are mutating fastest.

### 9.5 Per-corridor drift detection and model refresh (250, 260)

For each HS code × corridor pair, the system maintains the benchmark price distribution as a time series and applies stationarity tests of the KPSS/ADF type (Kwiatkowski, Phillips, Schmidt, & Shin, 1992) to detect distributional shifts indicative of technique mutation — launderers migrating from one goods category to another, or systematically adjusting invoice markups. A detected drift event triggers three actions: re-estimation of the affected benchmark distributions; an active-learning query burst in the affected corridor; and a model refresh incorporating the new labels. The loop is thereby self-driving — drift detection, targeted labelling, retraining, updated benchmarks — with no human in the re-engineering path. Human judgement remains where it belongs: in the investigations whose outcomes supply the labels. What the loop removes is the manual re-engineering between an outcome and the model's use of it.

10. Real-Time Scoring API

The scoring engine is exposed as a low-latency API (180). A subscribing institution — a bank, a payment processor, a trade-finance platform — submits a shipment's document set, or incremental document arrivals, and receives the risk score with reason codes within a latency budget suitable for pre-settlement interception; the disclosure states the target as sub-second for incremental updates on an existing graph neighbourhood. Two modes are supported. In synchronous scoring, the score is returned at payment time, for the decision the institution faces at that moment. In asynchronous monitoring, the score is recomputed as new documents arrive and subscribers are notified on threshold crossings — the mode that matches how trade documents actually appear, incrementally, over the life of a shipment, so that a shipment benign at invoice time can be caught at the arrival of a discrepant bill of lading. The precomputed benchmark distributions and incremental graph updates are what keep per-shipment scoring latency bounded independently of corpus size, and the reason codes returned with each score are, as Section 8 described, in the form case-management systems consume.

11. Full-Lifecycle E-Commerce Embodiment

A further embodiment extends the graph beyond business-to-business trade documents to direct-to-consumer (DTC) cross-border e-commerce, where the same laundering logic — value moved through fabricated or distorted commerce — operates across the full commercial lifecycle rather than across a documentary set. The graph is extended with six node types: parcel tracking records (carrier tracking numbers, scan events), delivery confirmations (proof-of-delivery records), advertising campaign records (spend, impressions, clicks per campaign, payee identity), web traffic sessions (site visits with attribution source), chargeback and dispute records, and customer complaint records.

The consistency-edge set is extended correspondingly, with five computationally verifiable assertion families spanning the lifecycle. Order-to-tracking existence: every paid order must resolve to at least one parcel tracking record within a plausible fulfilment window; absence is the e-commerce analogue of a phantom shipment. Tracking-to-delivery confirmation: tracking records must terminate in a delivery confirmation; systematic non-delivery amid completed payments indicates synthetic order flow. Advertising-spend-to-traffic consistency: reported advertising spend per campaign is cross-checked against independently observed traffic sessions attributed to that campaign; material spend without commensurate traffic — particularly where the payee is a related party of the advertiser — is scored as a fund-diversion indicator. Traffic-to-order conversion plausibility: the ratio of attributed traffic sessions to resulting orders is tested against category benchmarks; implausible conversion combined with suspect traffic sources indicates bot-generated orders. Chargeback-rate anomaly per entity: chargeback and dispute rates are computed per seller entity and tested against peer distributions, and complaint records are cross-checked against order volumes, where high volume with near-zero complaints indicates synthetic customers.

In this embodiment, chargeback outcomes (won or lost, fraud-coded) and complaint resolutions are ingested as labelled feedback into the same closed-loop self-iteration module of Section 9, so that evolving e-commerce fraud typologies — fake-order laundering, inflated marketing payments, friendly-fraud chargeback rings — are learned at the lifecycle-semantic level with the same drift-detection and active-learning machinery. The embodiment is the cross-validation idea restated for a setting in which the "documents" are lifecycle records: independent attestations, generated by independent parties (carriers, advertising platforms, card networks, customers), whose mutual consistency the operator of a synthetic storefront cannot fully control.

12. Alternative Embodiments

The disclosure records three variations. In one, the cross-validation graph is maintained as a federated structure across institutions, with consistency edges computed via privacy-preserving record linkage, so that no institution exposes customer data — the graph's cross-validation power being precisely what no single institution's partial documentary view can supply, the federated form addresses the obstacle to assembling it. In another, the benchmark price distributions are sourced from customs-published statistics rather than learned empirically. In a third, the scoring model is replaced by a rules engine over the consistency-edge scores, for jurisdictions requiring fully deterministic decision logic. The self-iteration loop operates identically in each embodiment — a statement about the architecture's centre of gravity: the loop and the graph are the invention's substance, and the scoring technology is a replaceable component within it.

13. Discussion and Limitations

Three technical effects are claimed for the disclosed architecture, each framed in the disclosure as measurable in controlled evaluation, and each stated here as a design claim rather than a result. First, false-positive reduction: documentary consistency evidence discriminates legitimate price dispersion from laundering markups that transaction statistics cannot separate, because the benchmark comparison is indexed by goods, corridor, and time, and is corroborated or undermined by the shipment's other documents. Second, adaptation without re-engineering: the drift-detection → active-learning → retraining loop updates documentary signatures automatically, so the system's adaptation latency to a novel typology is governed by label flow rather than by a manual feature-engineering cycle. Third, real-time operability: incremental graph updates and precomputed benchmark distributions bound per-shipment scoring latency independently of corpus size, which is what makes pre-settlement interception architecturally possible rather than aspirational. Whether each effect materialises at useful magnitude is an empirical question this paper does not answer.

The limitations are correspondingly plain. First and principally, nothing here is measured. This paper describes the system as disclosed in the patent filing. It reports no detection accuracy, no false-positive rates, no latency figures, and no deployment experience. The numerical values that appear — consistency scores in [0, 1], the sub-second target for incremental scoring updates — are the structural parameters and stated design targets of the disclosure, not experimental findings, and the typology feature definitions are stated functionally, at the level the disclosure states them.

Second, the system's vision is bounded by its documents. Cross-validation can only operate on documents that reach the graph; a shipment whose documents never enter the system presents no consistency surface, and the value of a subscribing institution's deployment grows with the breadth of document sources connected to it. The design mitigates — partial document sets still yield the consistency edges they support, and phantom-shipment features are expressly built from what is absent — but the boundary is real and is a property of the approach, not an implementation gap.

Third, the loop's quality is the quality of its labels. Feedback propagation, retraining with recency-weighted samples, and drift-triggered refresh all presume that investigation outcomes are accurate and that typology tags are consistently applied. A loop fed noisy or biased labels would learn structures that are not there, and would do so efficiently. The disclosure builds the machinery of adaptation; the governance of what the machinery learns from remains an institutional responsibility that the architecture assumes rather than discharges.

14. Conclusion

Transaction monitoring asks what a payment looks like; trade-based money laundering is designed to make that the wrong question. The system described here asks instead whether a shipment's documents agree with each other — whether the invoice, the bill of lading, the customs declaration, and the payment record describe one coherent movement of goods at a plausible price — and it builds that question into a graph in which the asking is computational: consistency edges with scores, from which typology features, calibrated risk, and documentary reason codes follow. Onto that structure it closes a loop in which investigations teach: confirmed structures reweight the graph, labels retrain the model, active learning directs scarce analyst attention to where uncertainty and information coincide, and per-corridor drift detection watches the benchmarks themselves for the signature of mutation. What is claimed is a relocation of the detection problem — from the payment, where TBML is invisible, to the documents, where it is conducted — together with the machinery that keeps the relocation current as the crime moves.

References

Blondel, V. D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment, 2008(10), P10008.

Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series with a unit root. Journal of the American Statistical Association, 74(366), 427–431.

FATF. (2006). Trade Based Money Laundering. Paris: Financial Action Task Force / OECD.

FATF & Egmont Group. (2020). Trade-Based Money Laundering: Trends and Developments. Paris: Financial Action Task Force.

Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis of stationarity against the alternative of a unit root: How sure are we that economic time series have a unit root? Journal of Econometrics, 54(1–3), 159–178.

Qu, S. (2026). Self-Iterating Detection of Trade-Based Money Laundering Using Multi-Source Trade Document Cross-Validation. U.S. Provisional Patent Application No. 64/170,275, filed October 6, 2026.

AI Disclosure: This working paper was prepared with AI-assisted writing and editing tools. The system description, arguments, and conclusions are the author's own, and the technical content is drawn from the author's patent disclosure cited above.