Source-linked AI summary
KONTOGRAPH: Verified Point-in-Time Feature Consistency and Amortised Explanation for Real-Time Anti-Money Laundering under a 200 ms Decision Budget
Ahmed Abolfadl
TL;DR
Instant euro payments remove AML batch and recovery windows, creating a need for rapid, defensible detection and explanation. KONTOGRAPH addresses this with a real-time temporal-graph pipeline, point-in-time feature compilation and measurement-focused evaluation. On simulated data, graph memory improved PR-AUC, while testing and deployment measurements exposed leakage defects and materially shifted alerts.
Problem
Instant payments remove overnight batch processing and settlement-based recovery, requiring AML scoring, explanation and decisions within the payment’s waiting period.
Method
KONTOGRAPH combines temporal graph learning, a pre-registered ablation protocol, point-in-time invariant testing and multi-backend feature execution for instant-payment AML.
Results
Per-node memory more than doubled PR-AUC from 0.0734 to 0.1717, with the rung 4 versus rung 1 difference +0.166 and 95% CI [0.105, 0.241].
Takeaways & Limitations
Measurement exposed three review-passing point-in-time defects and showed that a serving-format conversion shifted 12% of alerts while changing mean score by 7 × 10−8.
Takeaways & Limitations
All results use simulated payments; no transfer to production traffic is claimed, and the shared simulator-detector authorship leaves separability from real typologies uncertain.
Abstract
from arXiv · showhide
Regulation (EU) 2024/886 obliges European payment service providers to settle euro credit transfers in under ten seconds, around the clock. This removes both the overnight batch window in which anti-money-laundering (AML) analytics traditionally ran and the settlement delay that made recovery possible, forcing detection, explanation and decision inside a single-digit-second envelope. We present KONTOGRAPH, an end-to-end AML pipeline for the SEPA Instant rail built under a self-imposed 200 ms 99th-percentile budget, and report an empirical study on 1,562,860 simulated payments with injected typologies and deliberately incomplete labels. Three findings are of interest beyond the system itself. First, a temporal graph network with per-node memory improves PR-AUC over a gradient-boosted tabular baseline from 0.0053 to 0.1717, a paired day-blocked bootstrap difference of +0.166 with 95% CI [0.105, 0.241]; per-node memory alone more than doubles the score. Second, expressing each feature once and compiling it to three execution backends, with equivalence enforced by property-based tests that perturb the future, surfaced three point-in-time violations that code review had passed--each of which would have inflated reported performance. Third, and most consequential for practice, exporting the deployed tree ensemble to ONNX changed only $7.4 \times 10^{-8}$ in mean score yet altered 0.26% of decisions and inflated the alert volume by 12%, because 32-bit accumulation perturbs scores across a cost-optimal threshold of $3.98 \times 10^{-4}$. We argue that a serving-format conversion must be treated as a model change until measured, and that fidelity metrics for subgraph explainers can be vacuous when candidate neighbourhoods are small--a null result we report in full.
I. INTRODUCTION
KONTOGRAPH addresses instant-payment AML, where continuous settlement removes batch processing and recovery time while requiring rapid, defensible decisions. It combines temporal graph modelling with tested point-in-time features and evaluates the system on synthetic data.
- Regulation (EU) 2024/886 requires continuous SEPA Instant receipt and completion within ten seconds, eliminating AML’s overnight batch window and settlement lag.
- Scoring, explanation and decision must complete while the payer waits, with grounds that human investigators can verify.
- KONTOGRAPH treats point-in-time correctness as a tested invariant by compiling one feature specification to batch, streaming and online backends.
- The study ablates continuous-time graph structure and per-node memory on 1.5M simulated payments using paired day-blocked bootstrap intervals.
- The evaluation uses synthetic data with known ground truth, enabling measurement of label-selection bias but making no production-performance claim.
- Graph modelling is motivated by AML typologies whose defining fan-in, fan-out and chaining patterns are topological rather than attributes of individual transfers.
B. Explanation under a latency budget
The paper frames explanation as a real-time constraint: conventional perturbation and attribution methods are too slow, motivating amortised explanation and automated leakage testing. It also highlights that evaluation discipline matters under temporal and imbalanced data.
- Post-hoc graph explainers and SHAP require many model evaluations and are orders of magnitude beyond a 200 ms end-to-end budget.
- KONTOGRAPH pays perturbation-search cost offline, predicts explainer output in one forward pass, and uses counterfactual statements for investigator-facing narratives.
- The generated-dataset table is identified only by a caption and table label here, without passages specifying its row or column encodings.
- Precision-recall evaluation is motivated by severe class imbalance, while random cross-validation is invalid for temporal data.
- Property-based tests perturb future events and assert earlier computed values remain unchanged, without assuming a particular leakage mechanism.
III. SYSTEM ARCHITECTURE
KONTOGRAPH uses an agent-based simulation to create a deterministic German retail payment ecosystem with injected AML typologies and labels generated through a separate, biased confirmation process. Events move through schema-controlled streaming and lakehouse infrastructure.
- The simulator instantiates households, merchants, enterprises and banks whose payments follow an economic calendar and German holidays.
- Five AML typologies are injected with known ground truth, each having standard and evasive variants involving dwell time, fresh-account fan-out and sub-threshold amounts.
- Labels are generated separately from crime: confirmation requires an incumbent rules-engine flag and investigator concurrence, with verification latency distinct from event time.
- The dataset is a deterministic function of configuration, seed and code revision, with byte-identical regeneration checked by a determinism test.
- Events use Apache Avro and a schema registry, are published to a Kafka-compatible broker, and land in Apache Iceberg bronze, silver and gold layers with quality gates.
C. Feature platform
The feature platform specifies temporal semantics once and compiles them across offline, streaming and online execution. Agreement between implementations is tested, while serving keeps scoring synchronous and audit records decision-time state.
- Twenty-nine features are declared over entity, aggregation, window and filter, then lowered to a typed intermediate representation.
- Windows are left-closed and right-open, exclude the scored event and exclude events sharing its millisecond across all compiled targets.
- The intermediate representation compiles to DuckDB SQL, Flink SQL and an incremental online executor with per-entity accumulators.
- One specification and three compilers make training-serving agreement a tested anti-skew mechanism rather than a convention.
- Scoring runs synchronously because it is CPU-bound and single-digit milliseconds, while log persistence and alert publication remain asynchronous.
- Hash-chained decision logs capture model, feature, code, latency and explanation state at decision time, making tampering detectable and localisable.
IV. POINT-IN-TIME CORRECTNESS AS A TESTED INVARIANT
KONTOGRAPH treats point-in-time correctness as a machine-testable invariant rather than a coding convention. Property-based future perturbations and backend equivalence tests expose leakage and consistency defects that ordinary review can miss.
- A compiled feature must remain unchanged when events at or after its scoring time are altered.The invariant is tested without assuming a particular leakage mechanism.
- Property-based tests compare batch and online backends elementwise, failing the build on any divergence.
- Future-event perturbations detect leaks from window boundaries, joins, ordering errors, or future-looking enrichment through one mechanism-agnostic test.
- Three defects survived code review: fabricated first-event history, debtor-creditor key collisions, and simultaneous-event admission into strictly-prior windows.Each defect would have inflated reported performance.
- The evaluation protocol was committed before model development, including metrics, splits, cost model, ablations, and a planned negative result.The protocol was identified by hash 2eacb6e2.
B. Splits
The evaluation uses chronological, purged and embargoed splits with label-availability controls, while reporting ranking, capacity-constrained, and cost metrics. The held-out fold is extremely sparse, making day-level uncertainty important.
- Splits: Chronological splits use a one-day purge band and reject bands exceeding 25% of a fold.A 30-day band removed 79% of usable data and was deemed indefensible.
- Splits: Labels enter a fold only when their decided_ts precedes that fold’s availability cutoff.
- Splits: The held-out test fold contains 303,129 payments and 44 confirmed positives, a labelled positive rate of 1.45 × 10^-4.
- Metrics: PR-AUC is paired with precision and recall at 200 alerts per day and expected cost under an explicit investigation, friction, and prevention model.
- Metrics: Confidence intervals resample whole days because within-day payment dependence makes payment-level resampling too narrow.
- Operating point: GPU experiments use the 200-alert capacity threshold rather than fitting a test-fold threshold, avoiding leakage in reported cost figures.
VI. RESULTS
The ablation isolates graph structure and per-node memory, showing a large and statistically separated improvement over the baseline, while also demonstrating that ranking metrics can disagree with deployment cost. The system includes measured latency and explanation-serving considerations.
- Ablation ladder: +0.166 PR-AUC separates rung 4 from rung 1, with a 95% CI of [0.105, 0.241] excluding zero.
- Ablation ladder: Per-node memory more than doubles PR-AUC from 0.0734 to 0.1717.A separate run reproduced the direction and approximate magnitude, 0.0955 to 0.1555, as weak stability evidence.
- Cost metric: Rung 2 achieves nearly three times rung 1’s PR-AUC but costs EUR 53,529 versus EUR 20,459 and has lower recall, 0.795 versus 0.909, at capacity.
- Cost metric: Expected cost is recommended as the decision metric, with PR-AUC retained as a diagnostic because average ranking can mislead at the deployed threshold.
- Latency: The latency table reports per-stage decision timing for 2,000 payments on a single-threaded CPU.
- Explanations: The explanation table evaluates teacher–student performance over 57 instances.
C. Latency
KONTOGRAPH meets its latency budget with substantial headroom, but explanation results require careful interpretation because only alerting payments are explained and fidelity metrics are vacuous for tiny candidate sets.
- Latency interpretation: 98 of 2,000 payments are explained, making p99 rather than median explanation and counterfactual time the meaningful measure.Median times are zero because only alerting payments enter the explained path.
- Amortised explanation: 2.38 ms is the distilled student’s explanation time, versus 575.6 ms for the perturbation teacher.The teacher requires nearly three times the entire end-to-end budget, while the student remains inside it.
- Fidelity caveat: The reported fidelity values are not interpretable because the median explained instance has a single candidate edge.With one candidate, top-k Jaccard for k = 5 is 1.0 by construction and rank correlation is undefined for n < 2.
- Fidelity caveat: 0.904 and −0.038 principally measure candidate-set cardinality rather than student quality.The paper reports this as a null result rather than treating the figures as either success or substantive failure.
- Fidelity caveat: Explainer fidelity should be reported by candidate-count stratum or with a widened candidate set.The candidate-count distribution was not initially recorded, so the authors now record it and recommend publishing it alongside fidelity results.
VII. THREE FAILURE MODES FOUND BY MEASUREMENT
Measurement exposed failures in cost specification and serving-format conversion that ordinary evaluation could miss, including a material decision change after ONNX export.
- Inverted cost model: 2.7% recall was preferred over 97% recall because an initial cost model incorrectly credited recovered funds as revenue.The optimiser therefore treated alerting on almost nothing as optimal until the specification was corrected.
- Serving-format conversion: The ONNX export was compared with the original deployed gradient-boosted model across 303,129 test payments.The export was retained as a portability artifact, while the original artifact continued to serve decisions.
- Serving-format conversion: 3.98 × 10^-4 was the cost-optimal threshold at which 32-bit ONNX accumulation moved scores across the decision boundary.The source library accumulated in 64-bit, while the ONNX tree ensemble used 32-bit floating point.
- Serving-format conversion: A serving-format conversion is treated as a model change until parity is measured at the deployed threshold.Mean score error alone was judged insufficient for assessing decision fidelity.
- Serving-format conversion: 12% alert-volume inflation would have resulted from adopting the ONNX export at a capacity of 200 alerts per day.Published metrics would still have described a different model.
C. A generated report asserting absent evidence
Generated regulatory narratives can assert evidence absent from the underlying counterfactual, while the study’s synthetic, weakly labelled and non-deployed setting limits interpretation of its results.
- Generated report: “0 related payment(s)” incorrectly asserted transactional evidence for a feature-space counterfactual.Feature-space counterfactuals describe aggregate indicators and identify no transaction.
- Generated report: The generator now branches on counterfactual type and explicitly distinguishes aggregate indicators from identified transactions.This prevents feature-space explanations from being rendered as evidence about a specific payment.
- Data scope: All results use simulated payments, with no claim of transfer to production traffic.The simulator and detector share an author, and the injected typologies may be more separable than real ones.
- Label limitations: Only 14.5% of genuinely criminal payments receive confirmed labels, so precision figures are upper bounds on error under biased label selection.Evasive campaign variants are confirmed at roughly half the rate of standard variants.
- Label limitations: The test fold contains 44 confirmed positives, making bootstrap intervals correspondingly wide.The rung-4 versus rung-1 interval excludes zero, but adjacent middle-rung comparisons do not separate.
- Deployment scope: The latency budget measures only the gradient-boosted serving path; no latency claim is made for the graph model.Deploying the graph model would require threshold selection and latency measurement for a neural forward pass.
IX. CONCLUSION
KONTOGRAPH’s most transferable findings concern measurement: targeted tests exposed failures that plausible summary statistics and code review missed. The study also emphasizes reproducibility through regenerated artifacts and committed evaluation protocols.
- Measurement findings: Measurement tests exposed point-in-time violations, decision changes after model export, and vacuous explanation-fidelity metrics.These findings contradicted reasonable expectations that code review, negligible mean score error, or high overlap guaranteed correctness.
- Measurement findings: Comparing decisions rather than mean scores revealed that model export altered 12% of alerts despite changing mean score by 7 × 10−8.The discrepancy illustrates why deployment fidelity must be assessed at the decision level.
- Measurement findings: Recording the cardinality of compared neighbourhood sets exposed explanation-fidelity metrics that were vacuous at the observed size.A high Jaccard overlap alone was insufficient when candidate neighbourhoods were small.
- Measurement findings: A test that perturbs future events can reveal point-in-time leakage that ordinary code review misses.The paper reports three such defects, including fabricated first-event history, role-key collisions, and inclusion of simultaneous events.
- Reproducibility: Every reported figure can be regenerated from corresponding commands using available implementation, evaluation, decision-record, and generated artifacts.The evaluation protocol was committed before model development, and GPU metrics were independently recomputed to four decimal places.