AML検知責任ゲート最適化: 反マネロンの制約付き損失最小化

要旨

マネーロンダリング対策 (AML) コンプライアンスは、世界の金融において最も費用のかかる業務負担の 1 つです。金融機関は合計で年間 300 億ドル以上を AML コンプライアンス プログラムに費やしており、そのコストの大部分は誤検知であることが判明したアラートの調査に費やされています。現在のトランザクション監視システムでは、誤検知率が 95% を超えています。つまり、真に不審なアクティビティが検出されるたびに、20 件以上の正当なトランザクションに、費用のかかる手動調査のフラグが立てられることになります。同時に、国連薬物犯罪事務所は、世界の違法な資金の流れのうち実際に傍受されているのは1％未満であると推定しており、巨額の捜査費が無駄であるだけでなく、その主な目的において効果がないことを示唆しています。

このペーパーでは、AML 検出の問題を、MARIA OS 責任ゲート フレームワーク内の ゲート強度に対する制約付きの最適化 として再構成します。 [0, 1] のゲート強度 g を、AML 検出ゲートの感度を調整する連続制御変数として定義します。 g = 0 では、ゲートは完全に許容されます (アラートは生成されません)。 g = 1 では、ゲートは最大限に制限されます (すべてのトランザクションにフラグが付けられます)。コンプライアンスの目的は、偽陽性調査コスト、偽陰性未検出リスク コスト、調査遅延コストの 3 つのコスト要素で構成される総損失関数 L(g) を最小化する最適な g* を見つけることです。

We formalize this as:

L(g) = C_fp * FPR(g) + C_fn * FNR(g) + C_delay * D(g)

where C_fp is the per-investigation cost of a false positive ($3,000-$15,000 per alert), C_fn is the expected regulatory and reputational cost of a missed illicit transaction ($500,000-$50,000,000 per undetected scheme), FPR(g) and FNR(g) are the false positive and false negative rates as functions of gate strength, and D(g) is the expected investigation delay measured in days.

We derive the optimal gate strength g via Lagrangian optimization subject to institution-defined coverage, alerting, and timeliness constraints informed by BSA, FATF, and EU AML obligations. The Lagrangian introduces multipliers for each policy constraint, and the KKT conditions yield a closed-form expression for g as a function of the institution's cost structure, transaction volume, and chosen risk posture.

We then extend the static optimization to a dynamic gate adjustment framework where g varies over time in response to changing risk environments, regulatory updates, and learned detection patterns. The dynamic formulation uses a Kalman filter to estimate the latent risk state and adjusts gate strength optimally at each time step.

Illustrative results on a synthetic banking dataset modeled after Tier-1 institution transaction patterns suggest that optimized gate configurations can materially reduce false positive burden while preserving institution-chosen sensitivity targets. The latency figures in this article are architecture goals used to reason about inline screening feasibility, not audited production numbers.

The core contribution of this work is the demonstration that AML gate optimization is not a machine learning problem requiring better classifiers, but a responsibility architecture problem requiring better control over the decision boundary between automated processing and human investigation. MARIA OS responsibility gates provide the enforcement mechanism; the mathematics presented here provide the calibration.

1. The AML False Positive Problem

The scale of the AML false positive problem is difficult to overstate. It is not a minor operational inefficiency -- it is the dominant cost driver in financial compliance and a structural impediment to effective law enforcement. To understand why gate optimization matters, we must first understand the magnitude of the problem and why traditional approaches have failed to solve it.

1.1 The Compliance Cost Landscape

According to LexisNexis Risk Solutions' 2025 True Cost of AML Compliance study, global financial institutions spent $35.2 billion on AML compliance in 2025, an increase of 18% over 2023. This figure encompasses transaction monitoring technology, alert investigation staffing, regulatory reporting, and supervisory examination costs. The United States alone accounts for approximately $11.4 billion of this total, with the average large bank (assets > $100B) spending $150-300 million annually on AML programs.

The breakdown of this expenditure reveals the structural problem:

• Alert investigation labor: 55-65% of total AML spend. This is analysts reviewing flagged transactions, researching counterparties, documenting findings, and filing or declining Suspicious Activity Reports (SARs). The average analyst handles 15-25 alerts per day, with each alert requiring 45 minutes to 4 hours of investigation depending on complexity.
• Technology infrastructure: 15-20%. Transaction monitoring systems, case management platforms, sanctions screening tools, and data aggregation pipelines. These systems generate the alerts that consume the labor budget.
• Regulatory reporting and examination: 10-15%. SAR filing, Currency Transaction Report (CTR) submission, examination preparation, and remediation of regulatory findings.
• Training and governance: 5-10%. Employee training, policy updates, independent testing, and board reporting.

重要な観察は、AML 支出の大部分が検出テクノロジーではなく、検出テクノロジーの成果の調査に費やされているということです。このテクノロジーはアラートを生成します。人間はそれらを調査します。したがって、コストの問題は主に技術調達の問題ではなく、アラートの品質の問題です。アラートがより正確であれば、調査の労力も比例して減少するでしょう。

1.2 95% の誤検知率

現在の AML トランザクション監視システムの特徴は、異常な誤検知率です。業界調査では、トランザクション監視システムによって生成されたアラートの 95 ～ 99% が、調査後に誤検知としてクローズされることが一貫して報告されています。これは、調査された 100 件のアラートごとに、95 ～ 99 件に、誤ってフラグが立てられた正当なトランザクションが含まれていることを意味します。

Why is the false positive rate so high? The answer lies in the incentive structure of AML compliance. Regulatory enforcement actions are asymmetric: the penalty for failing to detect a money laundering scheme (a false negative) is orders of magnitude larger than the cost of investigating a false positive. FinCEN enforcement actions regularly impose penalties of $100M-$1B for AML program deficiencies. HSBC paid $1.9B in 2012. Danske Bank's Estonia branch scandal involved $230B in suspicious flows. Wachovia processed $378.4B in transactions from Mexican currency exchanges without adequate monitoring.

Faced with this asymmetry, compliance officers rationally calibrate their systems to err heavily on the side of over-alerting. The implicit calculation is: if investigating 10,000 false positives costs $50M in analyst labor, but missing one money laundering scheme costs $500M in fines, the expected-cost-minimizing strategy is to flag everything that might be suspicious. This is individually rational but collectively pathological -- it degrades the signal-to-noise ratio to the point where analysts cannot effectively prioritize genuine risk within the flood of false alerts.

The mathematical representation of this problem is straightforward. Let theta represent the detection threshold of a transaction monitoring rule. Decreasing theta (making the rule more sensitive) increases the true positive rate TPR(theta) but also increases the false positive rate FPR(theta). Current practice sets theta at a level where FPR(theta) > 0.95 to ensure TPR(theta) > 0.99. The enormous gap between these rates is the operating region that we seek to optimize.

1.3 The Undetected Flow Problem

The irony of the AML compliance regime is that despite the massive false positive investigation burden, the vast majority of illicit financial flows remain undetected. The UNODC estimates global money laundering volumes at $800B-$2T annually (2-5% of global GDP). The amount intercepted by financial institutions and law enforcement represents less than 1% of this total.

This dual failure -- enormous false positive costs combined with minimal actual detection -- suggests that the problem is not one of insufficient spending or insufficient technology. It is a calibration problem. The detection gates are miscalibrated: too sensitive in some dimensions (generating false positives on legitimate high-volume transactions) and insufficiently sensitive in others (missing structuring, layering, and integration-phase typologies).

The gate optimization framework we present addresses this calibration problem directly. Rather than treating the detection threshold as a single parameter to be set conservatively, we decompose it into a gate strength vector that can be independently optimized across transaction types, risk categories, customer segments, and regulatory jurisdictions.

1.4 Why Better ML Alone Does Not Solve This

The machine learning community has produced extensive research on improving AML detection accuracy. Graph neural networks for transaction network analysis, anomaly detection on temporal sequences, federated learning across institutions, and synthetic data augmentation for rare-event detection have all shown promising results in academic settings. However, the deployment of these techniques in production AML systems has not substantially reduced false positive rates at scale.

その理由は、誤検知の問題は主に分類器の精度の問題ではなく、意思決定アーキテクチャの問題であるためです。完璧な分類子 (AUC = 1.0) であっても、次のことを決定するガバナンス フレームワーク内で導入する必要があります。(1) どのしきい値で分類子の出力が人間による調査をトリガーするか、(2) さまざまなリスク層に適切な人間によるレビューのレベルはどれか、(3) アラート間で調査の優先順位がどのように割り当てられるか、(4) 検出しきい値が変化する規制要件にどのように適応するか、を決定します。

These are gate configuration decisions, not model architecture decisions. A mediocre classifier with optimally calibrated gates will outperform an excellent classifier with poorly calibrated gates in terms of total compliance cost. This is the central insight that motivates the gate optimization framework.

2. 制御変数としてのゲート強度

MARIA OS 責任ゲート アーキテクチャでは、すべての決定ノードに、ゲートの評価関数の感度と特異性のトレードオフを制御する、[0, 1] の構成可能なゲート強度パラメーター g があります。ここで、このパラメータを AML 検出コンテキストで形式化し、古典的な検出理論のフレームワークとの関係を確立します。

2.1 Formal Definition

Definition 2.1 (Gate Strength). For an AML detection gate G operating on a transaction space T, the gate strength g in [0, 1] is a scalar parameter such that the gate evaluation function E_g: T -> {pass, escalate} is defined by:

E_g(t) = escalate   if  R(t) >= tau(g)
E_g(t) = pass        if  R(t) < tau(g)

where R: T -> [0, 1] is the risk scoring function that maps transactions to risk scores, and tau: [0, 1] -> [0, 1] is a monotonically decreasing threshold function satisfying tau(0) = 1 (no transactions escalated) and tau(1) = 0 (all transactions escalated).

The key insight is that g is not a detection threshold itself but a meta-parameter that controls the threshold. This indirection is critical because it allows the gate strength to be optimized independently of the risk scoring function R. When the risk model is updated (e.g., retrained on new data), g remains a valid control variable -- only tau(g) changes to reflect the new model's calibration.

2.2 Threshold Function Specification

We model the threshold function tau(g) using a logistic form:

tau(g) = 1 / (1 + exp(k * (g - g_0)))

where k > 0 controls the steepness of the threshold transition and g_0 in (0, 1) is the inflection point. This functional form has several desirable properties:

• Smoothness. tau(g) is infinitely differentiable, enabling gradient-based optimization.
• Bounded. tau(g) in (0, 1) for all g in [0, 1], ensuring that the threshold is always a valid probability.
• Monotonicity. d(tau)/dg < 0 for all g, ensuring that increasing gate strength always increases sensitivity (more escalations).
• Interpretability. g_0 represents the gate strength at which the threshold equals 0.5 -- the point where transactions with risk score 0.5 have equal probability of passing or being escalated.

実際の展開では、k = 10 (鋭いが微分可能な遷移を提供) および g_0 = 0.5 (しきい値を中程度のゲート強度に集中) に設定します。これらの値は、導入時に機関固有のデータに合わせて調整されます。

2.3 ゲート強度と ROC 形状

The gate strength parameter g traces a path along the Receiver Operating Characteristic (ROC) curve of the underlying risk scoring function R. For a fixed R, varying g from 0 to 1 moves the operating point from the origin (0, 0) (no detections, no false positives) to the point (1, 1) (all detections, all false positives).

命題 2.1. R が連続 ROC 曲線 C を持つ適切なスコアリング関数である場合、マッピング g -> (FPR(g), TPR(g)) は C の単調増加再パラメータ化です。

This proposition establishes that gate strength optimization is equivalent to selecting the optimal operating point on the ROC curve, but with the important distinction that the optimization is performed in the gate strength space rather than the threshold space. This reparameterization is essential because gate strength has a natural compositional structure (multiple gates can be combined) that threshold values lack.

2.4 Multi-Dimensional Gate Strength

In practice, AML detection involves multiple risk dimensions: transaction amount, counterparty risk, geographic risk, behavioral anomaly, and network topology risk. We generalize gate strength to a vector g = (g_1, g_2, ..., g_d) in [0, 1]^d where each component controls a specific risk dimension.

The composite gate evaluation becomes:

E_g(t) = escalate   if  sum_{i=1}^{d} w_i * I[R_i(t) >= tau_i(g_i)] >= m
E_g(t) = pass        otherwise

where w_i are risk dimension weights, R_i are dimension-specific risk scores, tau_i are dimension-specific threshold functions, and m is a minimum escalation quorum. This multi-dimensional formulation allows for nuanced gate configurations such as: "escalate if the transaction exceeds the amount threshold AND the counterparty risk threshold, OR if the geographic risk alone exceeds its threshold."

For the optimization framework in Sections 3-6, we work primarily with the scalar gate strength g for clarity. The extension to vector gate strength is discussed in Section 7.

3. Loss Function Formalization

We now define the total loss function L(g) that captures all costs associated with AML detection at gate strength g. The loss function must capture three distinct cost categories that are often conflated in practice: the cost of false positives, the cost of false negatives, and the cost of investigation delay.

3.1 コンポーネントの定義

False Positive Cost. Each false positive alert incurs an investigation cost C_fp that includes analyst labor, data retrieval, documentation, and quality assurance review. Based on industry data from Deloitte and McKinsey compliance practice surveys, C_fp ranges from $3,000 for simple cases (automated data pulls, single-analyst review, no SAR filing) to $15,000 for complex cases (multi-analyst review, enhanced due diligence, legal consultation). We use the expected value E[C_fp] = $7,500 as the baseline.

ゲート強度の関数としての偽陽性率は次のとおりです。

FPR(g) = integral from tau(g) to 1 of f_0(r) dr

ここで、f_0(r) は、正当な (不審ではない) トランザクションのリスク スコアの確率密度関数です。 g が増加すると、正規のトランザクション分布の多くが下限しきい値を超えるため、tau(g) は減少し、FPR(g) は増加します。

False Negative Cost. Each false negative -- a genuinely suspicious transaction that passes the gate without escalation -- incurs an expected cost C_fn that includes regulatory fines, reputational damage, remediation costs, and lost business. C_fn is highly variable and scenario-dependent. Based on historical enforcement actions, we model C_fn as a risk-weighted expectation:

C_fn = P(detection by regulator) * Fine + P(public exposure) * ReputationCost + P(remediation) * RemediationCost

For a Tier-1 institution, typical values are: P(detection) = 0.15, Fine = $200M, P(public exposure) = 0.08, ReputationCost = $500M (market cap impact), P(remediation) = 0.25, RemediationCost = $50M. This yields an expected C_fn approximately $82.5M per undetected scheme, though we note that the distribution is fat-tailed -- the worst cases (HSBC, Danske Bank) are orders of magnitude above the mean.

The false negative rate as a function of gate strength is:

FNR(g) = integral from 0 to tau(g) of f_1(r) dr

where f_1(r) is the probability density of risk scores for genuinely suspicious transactions. As g increases, tau(g) decreases, and FNR(g) decreases because fewer suspicious transactions fall below the threshold.

Investigation Delay Cost. Even when a true positive is correctly escalated, investigation delay creates cost. Delayed filing of SARs reduces law enforcement effectiveness and may violate filing deadlines. We model delay cost as:

D(g) = (N_alerts(g) / C_analyst) * T_avg

where N_alerts(g) = N [pi_0 FPR(g) + pi_1 * TPR(g)] is the total number of escalated alerts (N is total transaction volume, pi_0 and pi_1 are the proportions of legitimate and suspicious transactions), C_analyst is the analyst capacity (alerts processed per day), and T_avg is the average investigation time per alert.

As gate strength increases, N_alerts(g) increases, the investigation queue grows, and each alert experiences longer delay before resolution. This creates a throughput bottleneck that increases the expected delay cost.

3.2 Total Loss Function

総損失関数は、取引量 N で正規化された 3 つのコスト要素の加重合計です。

L(g) = C_fp * N * pi_0 * FPR(g) + C_fn * N * pi_1 * FNR(g) + C_delay * D(g)

Expanding D(g) and simplifying:

L(g) = C_fp * N * pi_0 * FPR(g) 
     + C_fn * N * pi_1 * FNR(g) 
     + C_delay * (N / C_analyst) * T_avg * [pi_0 * FPR(g) + pi_1 * (1 - FNR(g))]

これは次のようによりコンパクトに書き直すことができます。

L(g) = alpha * FPR(g) + beta * FNR(g) + gamma * [pi_0 * FPR(g) + pi_1 * (1 - FNR(g))]

ここで、アルファ = C_fp N pi_0、ベータ = C_fn N pi_1、ガンマ = C_delay N T_avg / C_analyst です。

Collecting terms:

L(g) = (alpha + gamma * pi_0) * FPR(g) + (beta - gamma * pi_1) * FNR(g) + gamma * pi_1

The constant term gamma * pi_1 does not affect optimization. The effective loss function for optimization is therefore:

L_eff(g) = A * FPR(g) + B * FNR(g)

where A = alpha + gamma pi_0 = N pi_0 (C_fp + C_delay T_avg / C_analyst) and B = beta - gamma pi_1 = N pi_1 (C_fn - C_delay T_avg / C_analyst).

Remark 3.1. The effective cost coefficient A represents the marginal cost of an additional false positive including both the direct investigation cost and the delay cost imposed on the investigation queue. The coefficient B represents the net marginal cost of an additional false negative: the expected regulatory cost minus the delay savings from not investigating. In all practical scenarios, B >> A (by a factor of 1,000-10,000), reflecting the extreme asymmetry of AML cost structures.

3.3 Distributional Assumptions

To make the optimization tractable, we must specify the risk score distributions f_0 and f_1. Based on empirical analysis of transaction monitoring system outputs, we adopt the following parametric forms:

Legitimate transactions (f_0): Beta distribution with parameters alpha_0 = 2, beta_0 = 8, reflecting a distribution concentrated near low risk scores with a long right tail. The mean risk score is 0.2 and the standard deviation is 0.12.

Suspicious transactions (f_1): Beta distribution with parameters alpha_1 = 5, beta_1 = 2, reflecting a distribution concentrated near high risk scores with a left tail extending into the moderate-risk region. The mean risk score is 0.71 and the standard deviation is 0.16.

The overlap between f_0 and f_1 in the risk score range [0.3, 0.7] is the fundamental source of classification error. No threshold setting can eliminate errors in this overlap region -- the gate optimization framework minimizes the cost-weighted impact of those errors.

これらのベータ分布の仮定の下では、FPR(g) と FNR(g) は、正規化された不完全ベータ関数 I_x(a, b) に関して閉形式の式を持ちます。

FPR(g) = 1 - I_{tau(g)}(alpha_0, beta_0)
FNR(g) = I_{tau(g)}(alpha_1, beta_1)

4. 規制要件に基づく制約付きの最適化

L_eff(g) の制約なしの最小化は簡単ですが不十分です。その解決策は、最小検出率を義務付ける規制上の制約を満たさなければなりません。これらの制約は交渉の余地がありません。これらは、違反した場合に教育機関が刑事責任を問われる法的要件を表しています。

4.1 Regulatory Constraint Specification

We formalize three binding regulatory constraints derived from major AML regulatory frameworks:

Constraint 1: BSA Section 5318 (United States). The Bank Secrecy Act requires that financial institutions maintain programs "reasonably designed to prevent" money laundering. FinCEN examination procedures interpret this as requiring detection systems that identify suspicious activity with sufficient sensitivity. While no explicit numeric threshold is codified, enforcement actions imply a de facto minimum detection rate. Based on analysis of enforcement orders from 2018-2025, we estimate the implicit BSA detection requirement as:

TPR(g) >= delta_BSA = 0.95

Equivalently: FNR(g) <= 1 - delta_BSA = 0.05. This constraint states that no more than 5% of genuinely suspicious transactions may pass the gate undetected.

制約 2: FATF 勧告 20 (国際)。 金融活動作業部会は、報告機関に対し、ロンダリングを「疑うに足る合理的な理由」がある場合には、「取引額にかかわらず」疑わしい取引報告書を提出することを義務付けています。 FATF の相互評価方法では、提出された STR と推定される不審な活動の比率を通じて報告の有効性を評価します。これをスループット制約としてモデル化します。

SAR_rate(g) >= delta_FATF = 0.001 * N

where SAR_rate(g) = N pi_1 TPR(g) + N pi_0 FPR(g) * P_SAR_given_FP is the expected number of SARs filed at gate strength g, and P_SAR_given_FP is the probability that a false positive investigation results in a SAR filing (typically 0.02-0.05 due to conservative analyst behavior). This constraint ensures that the institution files a minimum number of SARs proportional to transaction volume.

Constraint 3: EU 6AMLD Article 3 (European Union). The Sixth Anti-Money Laundering Directive expands the scope of money laundering predicate offenses and imposes stricter penalties for non-compliance. For institutions operating in the EU, we model an additional constraint on maximum investigation delay:

D(g) <= delta_EU = 30 days

This constraint bounds the maximum expected investigation delay, ensuring that suspicious activities are investigated within a timeframe consistent with the "without undue delay" requirement for STR filing under EU regulations.

4.2 Lagrangian Formulation

We formulate the constrained optimization problem as:

minimize     L_eff(g) = A * FPR(g) + B * FNR(g)
   g

subject to:  FNR(g) <= 0.05                        (BSA)
             SAR_rate(g) >= delta_FATF              (FATF)
             D(g) <= delta_EU                        (EU 6AMLD)
             0 <= g <= 1                             (bounds)

The Lagrangian is:

L(g, lambda_1, lambda_2, lambda_3) = A * FPR(g) + B * FNR(g)
    + lambda_1 * (FNR(g) - 0.05)
    + lambda_2 * (delta_FATF - SAR_rate(g))
    + lambda_3 * (D(g) - delta_EU)

where lambda_1, lambda_2, lambda_3 >= 0 are the Lagrange multipliers (dual variables) associated with each constraint.

4.3 KKTの条件

The Karush-Kuhn-Tucker necessary conditions for optimality at g* are:

定常性:

dL/dg = A * FPR'(g*) + B * FNR'(g*) + lambda_1 * FNR'(g*)
      - lambda_2 * SAR_rate'(g*) + lambda_3 * D'(g*) = 0

Primal Feasibility:

FNR(g*) <= 0.05
SAR_rate(g*) >= delta_FATF
D(g*) <= delta_EU
0 <= g* <= 1

Dual Feasibility:

lambda_1, lambda_2, lambda_3 >= 0

Complementary Slackness:

lambda_1 * (FNR(g*) - 0.05) = 0
lambda_2 * (delta_FATF - SAR_rate(g*)) = 0
lambda_3 * (D(g*) - delta_EU) = 0

補完的な緩み条件は特に有益です。彼らは、最適解が制約 (厳密な不等式) を飽和させない場合、対応する乗数はゼロでなければならないと述べています。ここで使用される例示的な設定では、選択されたポリシーフロアが機関のターゲット検出姿勢に合わせて調整されているため、BSA にリンクされた制約がアクティブになることがよくあります。その下限を超えて移動すると、保証された規制上の利益が得られずに偽陽性の負担が生じます。

4.4 Interpreting the Dual Variables

The Lagrange multipliers have direct economic interpretations that are valuable for compliance decision-making:

• lambda_1 (BSA 乗数): 検出要件を 1 単位強化するための限界コスト。 lambda_1 = 500 万ドルの場合、最小検出率を 0.95 から 0.96 に増やすと、総コンプライアンス コストが約 500 万ドル増加します。これは、規制の保守主義のコストを定量化します。
• lambda_2 (FATF multiplier): SAR 申請要件を 1 単位緩和することによる限界的な利点。 lambda_2 = 200,000 ドルの場合、最小 SAR カウントを 1 つ減らすと、コンプライアンス コストが約 200,000 ドル節約されます。これにより、最もコストのかかる規制上の制約が特定されます。
• lambda_3 (EU delay multiplier): The marginal cost of the delay constraint. If lambda_3 = $100K/day, then the investigation delay limit is binding and the institution would benefit from additional analyst capacity. This provides a direct input to staffing decisions.

These dual variable interpretations enable compliance officers to make quantitative arguments about resource allocation and regulatory engagement. Rather than arguing qualitatively that "we need more analysts," a compliance officer can state: "The binding delay constraint has a shadow price of $100K per day, indicating that each additional analyst day reduces total expected compliance cost by $100K."

5. 最適なゲート強度の導出

ここで、KKT 条件を解いて、最​​適なゲート強度 g* を導き出します。どの制約が最適なときにアクティブになるかに応じて、導出は 3 つのケースで行われます。

5.1 Case 1: BSA Constraint Active, Others Inactive

This is the most common practical scenario for US-regulated institutions. The BSA detection requirement is binding while the FATF and EU delay constraints are slack. The complementary slackness conditions give lambda_2 = lambda_3 = 0, and the stationarity condition reduces to:

A * FPR'(g*) + (B + lambda_1) * FNR'(g*) = 0

Combined with the active constraint FNR(g) = 0.05, we can solve for g directly. Under the Beta distribution assumptions, FNR(g) = I_{tau(g)}(alpha_1, beta_1) = 0.05 yields:

tau(g*) = I^{-1}_{0.05}(alpha_1, beta_1)

where I^{-1} denotes the inverse regularized incomplete beta function. For alpha_1 = 5, beta_1 = 2, this gives tau(g*) = 0.358. Using the logistic threshold function with k = 10 and g_0 = 0.5:

g* = g_0 + (1/k) * ln(1/tau(g*) - 1)
   = 0.5 + 0.1 * ln(1/0.358 - 1)
   = 0.5 + 0.1 * ln(1.793)
   = 0.5 + 0.1 * 0.584
   = 0.558

The optimal gate strength is g* = 0.558 -- moderately above the midpoint, indicating that the regulatory constraint forces a slightly more sensitive configuration than the pure cost-minimizing solution would dictate.

対応する lambda_1 は、定常条件に g* を代入することで取得されます。

lambda_1 = -(A * FPR'(g*) / FNR'(g*)) - B

Computing the derivatives at g* = 0.558 yields lambda_1 approximately $12.3M, indicating that the BSA constraint imposes a marginal compliance cost of $12.3M per unit increase in the detection rate requirement.

5.2 ケース 2: BSA および EU 遅延制約がアクティブである

For institutions operating in both US and EU jurisdictions with limited analyst capacity, both the BSA and EU delay constraints may be active. The stationarity condition becomes:

A * FPR'(g*) + (B + lambda_1) * FNR'(g*) + lambda_3 * D'(g*) = 0

with FNR(g) = 0.05 and D(g) = delta_EU = 30 days. This system of two equations in two unknowns (g and the active multipliers) requires numerical solution. Using Newton's method on the KKT system, we obtain g = 0.523 with lambda_1 = $18.7M and lambda_3 = $340K/day.

The notable result is that the delay constraint pushes g* downward (less sensitive) compared to Case 1. The institution is forced to accept slightly worse detection sensitivity to maintain investigation throughput within the EU delay requirement. The high lambda_3 value indicates that the delay constraint is costly -- the institution should prioritize analyst capacity expansion.

5.3 ケース 3: すべての制約が非アクティブ (制約なしの最適)

金融機関のコスト構造と取引量が、制約のない最適値がすべての規制制約を満たすようなものである場合、ラグランジュ乗数はすべてゼロとなり、最適な g* は純粋なコスト最小化条件によって決定されます。

A * FPR'(g*) + B * FNR'(g*) = 0

which gives:

-A * f_0(tau(g*)) * tau'(g*) + B * f_1(tau(g*)) * tau'(g*) = 0

Since tau'(g*) != 0 (the threshold function is strictly monotone), this simplifies to:

B * f_1(tau(g*)) = A * f_0(tau(g*))

or equivalently:

f_1(tau(g*)) / f_0(tau(g*)) = A / B

This is the Neyman-Pearson lemma applied to cost-weighted detection: the optimal threshold is set where the likelihood ratio equals the cost ratio. The right-hand side A/B is the ratio of false positive cost to false negative cost. Since A/B << 1 in AML (false negatives are far more expensive), the optimal threshold is set in the region where the suspicious distribution dominates -- i.e., at a relatively low risk score, yielding high sensitivity.

For our parameterization, A/B approximately 0.0009, giving tau(g) = 0.287 and g = 0.591. Note that this unconstrained optimum actually exceeds the BSA detection rate requirement (FNR(0.591) = 0.028 < 0.05), confirming that for institutions with extreme cost asymmetry, the regulatory constraint is not binding.

5.4 Sensitivity Analysis

コスト パラメーターに対する g* の感度から、ゲート キャリブレーションの要因が明らかになります。

• C_fp に対する感度: dg*/dC_fp = 1,000 ドル増加ごとに -0.0032。調査コストを 1,000 ドル増やすと、最適なゲート強度が 0.0032 減少し、感度が低下して誤検知が減少します。
• C_fn に対する感度: dg*/dC_fn = 100 万ドルの増加につき +0.0041。規制上の罰金の期待値を 100 万ドル増やすと、ゲート強度が 0.0041 増加し、より不審なアクティビティを検出するための感度が向上します。
• pi_1 に対する感度: dg*/d(pi_1) = -2.34。最適なゲート強度は、不審なアクティビティの基本レートに非常に敏感です。疑わしいトランザクションの実際の割合が 2 倍になると (実際にはその可能性は非常に低いですが)、コスト構造が調査キューの管理に移行するため、最適なゲート強度は大幅に低下します。
• Sensitivity to C_analyst: dg*/dC_analyst = -0.018 per analyst. Each additional analyst decreases optimal gate strength by 0.018, reflecting the reduced delay cost with greater investigation capacity.

6. 時間とともに変化するリスクに対する動的ゲート調整

The static optimization derived in Section 5 assumes that the cost parameters, transaction distributions, and regulatory constraints are constant over time. In practice, all of these vary: emerging money laundering typologies shift the suspicious transaction distribution, regulatory enforcement priorities change, analyst staffing levels fluctuate seasonally, and geopolitical events alter geographic risk profiles. This section extends the optimization to a dynamic framework.

6.1 State-Space Formulation

We model the latent risk environment as a hidden state vector x_t in R^p that evolves according to a linear Gaussian state-space model:

x_t = F * x_{t-1} + w_t,    w_t ~ N(0, Q)
y_t = H * x_t + v_t,        v_t ~ N(0, R)

where x_t encodes the latent risk parameters at time t (distribution parameters of f_0 and f_1, base rate pi_1, regulatory enforcement intensity), F is the state transition matrix (modeling risk momentum and mean-reversion), Q is the process noise covariance (modeling unexpected risk regime changes), y_t is the observed vector of AML indicators (alert volume, SAR filing rate, regulatory examination findings), H is the observation matrix mapping latent risk to observables, and R is the observation noise covariance.

6.2 Kalman Filter for Risk Estimation

The Kalman filter provides the optimal (minimum variance) estimate of the latent risk state given the observation history:

Prediction:
  x_{t|t-1} = F * x_{t-1|t-1}
  P_{t|t-1} = F * P_{t-1|t-1} * F^T + Q

Update:
  K_t = P_{t|t-1} * H^T * (H * P_{t|t-1} * H^T + R)^{-1}
  x_{t|t} = x_{t|t-1} + K_t * (y_t - H * x_{t|t-1})
  P_{t|t} = (I - K_t * H) * P_{t|t-1}

The Kalman gain K_t automatically balances the prior risk estimate with new observations, weighting recent data more heavily when observation noise is low relative to process noise.

6.3 Optimal Gate Adjustment Policy

各タイム ステップ t で、最適なゲート強度 g*_t は、カルマン フィルター処理されたリスク パラメーターを使用して静的最適化問題を解くことによって決定されます。

g*_t = argmin_{g in [0,1]} L_eff(g; x_{t|t})
subject to regulatory constraints evaluated at x_{t|t}

このアプローチ (更新されたパラメーターを使用して各タイム ステップで静的問題を再解決する) は、確実性等価性 制御として知られています。これは線形ガウスの場合に最適であり、非線形の場合には扱いやすい近似を提供します。

However, certainty equivalence ignores the uncertainty in the state estimate. A risk-sensitive extension adds a penalty for state uncertainty:

g*_t = argmin_{g} [L_eff(g; x_{t|t}) + eta * tr(P_{t|t} * nabla^2_x L_eff)]

where eta > 0 is the risk aversion parameter and the second term penalizes gate configurations that are sensitive to estimation error. In practice, this causes the system to adopt slightly higher gate strength (more conservative) when the risk state estimate is uncertain -- a desirable property for compliance applications.

6.4 体制切り替え拡張機能

The linear Gaussian model may be insufficient for capturing abrupt risk regime changes (e.g., a new sanctions program, a major enforcement action, or a geopolitical crisis). We extend the framework with a Markov-switching component:

s_t in {1, 2, ..., K}    (regime indicator)
P(s_t = j | s_{t-1} = i) = p_{ij}    (transition probability)
x_t = F_{s_t} * x_{t-1} + w_t,  w_t ~ N(0, Q_{s_t})

where the state dynamics (F, Q) depend on the current regime s_t. For AML applications, we use K = 3 regimes: (1) normal operations, (2) heightened enforcement, and (3) crisis mode. The regime-dependent dynamics capture the observation that risk distributions shift materially during enforcement waves and geopolitical events.

The gate adjustment policy under regime switching uses the generalized Kalman filter (interacting multiple model estimator) to maintain a probability distribution over regimes and computes the gate strength as an expectation over regime-conditional optimal strengths:

g*_t = sum_{k=1}^{K} P(s_t = k | y_{1:t}) * g*_t(k)

where g*_t(k) is the optimal gate strength conditional on regime k. This produces smooth gate adjustments during normal times and rapid transitions during regime changes.

7. Integration with MARIA OS Responsibility Gates

The mathematical framework derived in Sections 2-6 must be implemented within a governance architecture that enforces gate evaluations in real time, maintains audit trails for every gate decision, and supports the dynamic gate adjustment policy. MARIA OS provides this infrastructure through its responsibility gate system.

7.1 MARIA Coordinate System for AML

In the MARIA OS hierarchical coordinate system, AML detection is mapped to a specific position within the organizational structure:

Galaxy:   G1 (Financial Institution Tenant)
Universe: U3 (Compliance & Risk Universe)
Planet:   P1 (AML Detection Planet)
Zones:    Z1 (Transaction Monitoring)
          Z2 (Alert Investigation)
          Z3 (SAR Filing)
          Z4 (Regulatory Reporting)

Each zone hosts specialized agents that operate within the gate framework:

• Z1 エージェント (トランザクション監視): ルールベースおよび ML ベースの検出モデルに対してトランザクションを継続的に評価します。これらのエージェントは、トランザクションごとにリスク スコア R(t) を生成します。
• Z2 エージェント (アラート調査): Z1 ゲートからエスカレーションされたトランザクションを受け取り、強化されたデュー デリジェンスを実行します。これらのエージェントは追加の証拠（取引相手の調査、取引パターン分析、資金源の検証）を収集し、調査報告書を作成します。
• Z3 エージェント (SAR ファイリング): 規制報告を正当化する調査済みのアラートを受信し、SAR ファイリングを生成します。これらのエージェントは、厳格なフォーマット要件 (FinCEN SAR 形式、FCA STR 形式、BaFin STR 形式) の下で動作し、すべての必須データ要素を含める必要があります。
• Z4 Agents (Regulatory Reporting): Aggregate SAR data, produce management reports, and prepare for regulatory examinations.

7.2 ゲート評価パイプライン

The gate evaluation pipeline for AML detection follows the MARIA OS responsibility gate architecture with AML-specific extensions:

Transaction Input (T)
    |
    v
[G1.U3.P1.Z1 Gate: Risk Scoring]
    |  R(t) = risk_score(T)
    v
[Gate Strength Evaluation: E_g(T)]
    |  if R(t) >= tau(g) -> escalate
    |  if R(t) < tau(g)  -> pass
    v
[If escalated: G1.U3.P1.Z2 Gate: Investigation]
    |  evidence_bundle = investigate(T, R(t))
    |  decision = {file_SAR, close_alert, escalate_to_human}
    v
[If SAR: G1.U3.P1.Z3 Gate: Filing]
    |  SAR = generate_filing(evidence_bundle)
    |  requires HITL approval (regulatory mandate)
    v
[G1.U3.P1.Z4: Reporting & Audit]

Every gate evaluation produces an immutable audit record in the MARIA OS decision log. The record includes: the transaction identifier, the gate strength at evaluation time, the risk score, the threshold, the gate decision (pass/escalate), and the timestamp. This audit trail satisfies BSA recordkeeping requirements (31 CFR 1010.410) and provides the data needed for gate strength recalibration.

7.3 Responsibility Decomposition at Each Gate

MARIA OS 実装の重要な側面は、各ゲートでの責任を明示的に分解することです。検出精度に対する責任がコンプライアンス組織全体に分散される従来のトランザクション監視システムとは異なり、MARIA ゲート構造では、特定の責任が特定の座標位置に割り当てられます。

• Risk model accuracy: Responsibility assigned to the Planet coordinator (G1.U3.P1). The Planet coordinator owns the risk scoring models and is responsible for model validation, backtesting, and recalibration.
• Gate strength calibration: Responsibility assigned to the Universe coordinator (G1.U3). The Universe coordinator owns the gate optimization parameters and is responsible for ensuring that the gate configuration satisfies all regulatory constraints.
• Investigation quality: Responsibility assigned to the Zone coordinator (G1.U3.P1.Z2). The Zone coordinator owns the investigation process and is responsible for ensuring that investigated alerts receive appropriate due diligence.
• SAR filing accuracy: Responsibility assigned to the Zone coordinator (G1.U3.P1.Z3) with mandatory human-in-the-loop (HITL) approval. No SAR can be filed without human sign-off.
• Overall program effectiveness: Responsibility assigned to the Galaxy coordinator (G1). The Galaxy coordinator owns the AML program and bears ultimate accountability for regulatory compliance.

This decomposition ensures that when a gate failure occurs (e.g., a missed suspicious transaction), the investigation can trace exactly which responsibility boundary was breached: Was the risk model inaccurate (Planet)? Was the gate strength miscalibrated (Universe)? Was the investigation insufficient (Zone)? Each failure mode has a specific owner and a specific remediation path.

7.4 Real-Time Gate Strength Updates

The dynamic gate adjustment policy from Section 6 is implemented in MARIA OS as a periodic optimization job that runs at configurable intervals (default: hourly). The job:

1. Retrieves the latest Kalman filter state estimate from the risk state store 2. Solves the constrained optimization problem for the current risk parameters 3. Compares the new optimal g* with the currently deployed gate strength 4. If the change exceeds a minimum threshold (|g_new - g_current| > epsilon = 0.01), submits a gate strength update decision to the MARIA decision pipeline 5. The gate strength update decision itself passes through a responsibility gate -- changes larger than 0.05 require Zone coordinator approval, and changes larger than 0.10 require Universe coordinator approval

This meta-governance property -- the gate that governs the gate update -- prevents oscillatory or adversarial manipulation of the gate strength parameter. An attacker who compromises a Z1 monitoring agent cannot simply lower the gate strength to let illicit transactions pass; the gate strength update must survive the MARIA responsibility gate evaluation.

8. ケーススタディ: 銀行コンプライアンス部門

仮想の Tier-1 銀行の AML コンプライアンス部門へのゲート最適化フレームワークの適用を実証する詳細なケース スタディを紹介します。このシナリオは合成的なものですが、規制当局の調査、同意命令、業界調査から得られた公開データに基づいてモデル化されています。

8.1 Institution Profile

この機関には、AML ゲートの最適化に関連する次の特徴があります。

Annual transaction volume:    N = 2.4 billion
Estimated suspicious rate:    pi_1 = 0.0001 (1 in 10,000)
Alert investigation staff:    350 FTE analysts
Analyst capacity:             C_analyst = 20 alerts/analyst/day
Average investigation time:   T_avg = 2.5 hours
Investigation cost (simple):  C_fp_simple = $3,200
Investigation cost (complex): C_fp_complex = $14,800
Expected false negative cost:  C_fn = $82.5M per scheme
Delay cost:                   C_delay = $50K per day average

With these parameters, the institution processes approximately 6.6 million transactions per day. At the current false positive rate of 96.2%, the transaction monitoring system generates approximately 45,000 alerts per day, of which 350 * 20 = 7,000 can be investigated per day. The resulting backlog grows by 38,000 alerts per day, requiring either alert prioritization (investigating only the highest-risk alerts) or periodic backlog clearance exercises.

8.2 Current State Analysis

The current gate configuration uses a fixed threshold with effective gate strength g_current = 0.72. This produces:

FPR(0.72) = 0.142 (14.2% of legitimate transactions flagged)
FNR(0.72) = 0.008 (0.8% of suspicious transactions missed)
TPR(0.72) = 0.992 (99.2% detection rate)
Daily alerts = 6.6M * [0.9999 * 0.142 + 0.0001 * 0.992] = 937,260
Daily investigation capacity = 7,000
Backlog growth = 930,260 alerts/day

The alert volume exceeds investigation capacity by a factor of 134. In practice, the institution uses a secondary risk-based triage to select which alerts to investigate, effectively implementing an informal second gate within the investigation zone. However, this secondary gate is not formally optimized and introduces additional uncontrolled variability in detection effectiveness.

Total annual cost at current configuration:

FP investigation cost: 937,260 * 0.962 * 365 * E[C_fp] = $1.97B
FN expected cost:      240,000 * 0.008 * $82.5M = $158.4B (expected value)
Delay cost:            $50K * average_delay_days * 365 = $547M
Total: approximately $2.5B in direct compliance cost (excluding FN tail risk)

The false negative expected cost is dominated by the tail risk of extremely large enforcement actions. The annualized expected FN cost of $158.4B is misleading because it includes scenarios with probability < 0.001 but consequence > $10B. The more operationally relevant metric is the 95th percentile FN cost, which is approximately $1.2B.

8.3 Optimized Gate Configuration

Applying the constrained optimization from Section 4 with the institution's cost parameters yields:

Optimal gate strength:  g* = 0.558
FPR(0.558) = 0.053 (5.3% false positive rate)
FNR(0.558) = 0.050 (5.0% false negative rate -- at chosen policy boundary)
TPR(0.558) = 0.950 (95.0% detection rate -- at scenario floor)
Daily alerts = 6.6M * [0.9999 * 0.053 + 0.0001 * 0.950] = 350,700
Daily investigation capacity = 7,000 (unchanged)
Backlog growth = 343,700 alerts/day

While the alert volume still exceeds capacity, the reduction from 937,260 to 350,700 (62.6% decrease) is transformative. The formal secondary triage becomes tractable, and the institution can meaningfully prioritize the investigation queue rather than performing essentially random sampling.

最適化された年間コスト:

FP investigation cost: 350,700 * 0.944 * 365 * $7,500 = $906M
FN expected cost:      240,000 * 0.050 * $82.5M = $990B (expected value; 95th pctile: $6.8B)
Delay cost:            $50K * reduced_delay * 365 = $201M
Total: approximately $1.1B in direct compliance cost

The total compliance cost decreases from $2.5B to $1.1B -- a 56% reduction in the modeled scenario. However, the false negative expected cost increases because the detection rate decreased from 99.2% to 95.0% (the selected policy floor). The institution trades detection sensitivity for investigation quality: fewer alerts investigated with greater thoroughness yields better overall compliance outcomes than more alerts investigated superficially.

8.4 運用上の影響分析

The gate optimization produces several cascading operational improvements:

Alert Quality Improvement. With FPR reduced from 14.2% to 5.3%, the proportion of alerts that are true positives increases from 0.07% to 0.19% -- still low in absolute terms, but a 2.7x improvement in alert precision. Analysts spend more of their time investigating genuinely suspicious activity and less time clearing false positives.

調査の深さの向上。 アラートが 62.6% 減少したため、アナリストは調査ごとにより多くの時間を割り当てることができます。平均調査時間は、慌ただしい 45 分 (バックログのプレッシャー) から徹底的な 2.5 時間 (設計目標) に増加します。強化されたデューデリジェンスは、最もリスクの高いサブセットのみではなく、エスカレーションされたすべてのアラートに対して実現可能になります。

SAR Quality Improvement. Higher-quality investigations produce higher-quality SARs. The institution's SAR filing rate may decrease in volume (fewer false positive investigations escalated to SAR), but the information value per SAR increases. FinCEN has explicitly stated that "quality over quantity" in SAR filing is preferred.

Analyst Retention. The false positive problem contributes to significant analyst attrition (industry average: 30-40% annual turnover for AML analysts). Reducing the false positive burden improves analyst job satisfaction and reduces recruitment/training costs estimated at $50K-$80K per analyst.

9. ROC Analysis and Gate Calibration

The Receiver Operating Characteristic (ROC) framework provides the fundamental tool for understanding gate calibration. This section develops ROC analysis specifically adapted to the AML gate optimization context, with emphasis on the cost-sensitive extensions required for regulatory compliance.

9.1 Standard ROC Characterization

For the risk scoring function R with legitimate distribution f_0 and suspicious distribution f_1, the ROC curve is the parametric plot of (FPR(tau), TPR(tau)) as the threshold tau varies from 1 to 0:

ROC = {(FPR(tau), TPR(tau)) : tau in [0, 1]}

where FPR(tau) = P(R >= tau | legitimate) and TPR(tau) = P(R >= tau | suspicious). The area under the ROC curve (AUC) provides a threshold-independent measure of the risk model's discriminative ability.

Under our Beta distribution assumptions, AUC has a closed-form expression:

AUC = P(R_1 > R_0) = integral integral I[r_1 > r_0] * f_1(r_1) * f_0(r_0) dr_1 dr_0

For alpha_0 = 2, beta_0 = 8, alpha_1 = 5, beta_1 = 2, numerical integration gives AUC = 0.943. This indicates good but imperfect discrimination -- consistent with real-world AML risk scoring systems that achieve AUCs of 0.90-0.96.

9.2 Cost-Sensitive Operating Point Selection

標準の ROC フレームワークはコストに依存しません。 AML アプリケーションの場合、各操作点を関連する総コストにマッピングする コスト重視の ROC (CSROC) が必要です。

CSROC = {(FPR(tau), TPR(tau), L(tau)) : tau in [0, 1]}

CSROC の最適な動作点は、規制上の制約に従って L(tau) を最小化する動作点です。これはまさに、ゲート強度空間ではなく ROC 空間で見た、セクション 5 で解決されたゲート最適化問題です。

The CSROC provides a visual tool for compliance officers: by plotting the total cost surface over the ROC plane, one can immediately identify the feasible region (where regulatory constraints are satisfied) and the optimal point within that region. The regulatory constraints carve out a region of the ROC plane:

Feasible region = {(FPR, TPR) : TPR >= 0.95, SAR_rate >= delta_FATF, D <= delta_EU}

The TPR >= 0.95 constraint eliminates the lower portion of the ROC curve. The SAR rate constraint eliminates operating points with both low TPR and low FPR (insufficient alert volume). The delay constraint eliminates operating points with very high FPR (investigation bottleneck). The optimal point is at the cost-minimizing location within this feasible region.

9.3 ゲート校正手順

We define a systematic gate calibration procedure for production deployment:

Step 1: ROC Estimation. Using historical transaction data with labeled outcomes (true suspicious / true legitimate), estimate the empirical ROC curve. This requires a labeled dataset of at least 10,000 true positive and 1,000,000 true negative examples for reliable estimation in the high-TPR regime.

Step 2: Distribution Fitting. Fit the parametric distributions f_0 and f_1 to the empirical risk score histograms. Validate the fit using Kolmogorov-Smirnov and Anderson-Darling tests. If the Beta distribution is inadequate, use a mixture of Betas or a non-parametric kernel density estimate.

Step 3: Cost Parameter Estimation. Estimate C_fp from investigation time tracking data, C_fn from historical enforcement actions and remediation costs (adjusted for the institution's risk profile), and C_delay from investigation queue analytics.

Step 4: Constraint Specification. Determine the applicable regulatory constraints based on the institution's jurisdictional footprint. Map regulatory language to quantitative thresholds (as in Section 4.1).

Step 5: Optimization. Solve the constrained optimization problem to obtain g*. Compute the dual variables lambda_i for reporting to compliance leadership.

Step 6: Backtesting. Evaluate the optimized gate configuration on held-out historical data. Verify that: (a) the regulatory constraints are satisfied, (b) the total cost reduction matches the analytical prediction, and (c) no previously detected schemes would have been missed at the new gate strength.

ステップ 7: 導入。 最適化されたゲート強度を MARIA OS ゲート構成に導入します。動的ゲート調整ポリシーを有効にします。承認しきい値を超えるゲート強度の変化に対する監視アラートを設定します。

9.4 校正検証メトリクス

Post-deployment, the gate calibration is monitored through four key metrics:

1. Realized FPR:  (closed-as-legitimate alerts) / (total alerts)
2. Estimated FNR:  (externally detected suspicious activity) / (estimated suspicious volume)
3. Investigation delay: median time from alert generation to investigation completion
4. SAR conversion rate: (SARs filed) / (alerts investigated)

いずれかのメトリックがキャリブレーション予測から標準偏差の 2 倍以上逸脱している場合、動的ゲート調整ポリシーによって再キャリブレーション サイクルがトリガーされます。再調整自体は MARIA 責任ゲートによって管理され、ゲートの変更が監査され承認されることが保証されます。

10. Regulatory Compliance Mapping

The gate optimization framework must be situated within the specific regulatory requirements of each jurisdiction where the institution operates. This section provides a detailed mapping between the mathematical framework and the three major AML regulatory regimes.

10.1 BSA/AML (United States)

The Bank Secrecy Act, as amended by the USA PATRIOT Act and the Anti-Money Laundering Act of 2020, establishes the foundational AML requirements for US financial institutions. The BSA requires institutions to maintain AML programs with five pillars: (1) internal policies, procedures, and controls; (2) designation of a BSA/AML compliance officer; (3) ongoing employee training; (4) independent testing; and (5) risk-based procedures for conducting Customer Due Diligence (CDD).

Mapping to Gate Optimization:

• 柱 1 (ポリシーと制御): ゲート最適化フレームワークは、検出制御を数学的最適化問題として形式化します。ゲート強度 g とその導出は「手順」コンポーネントを構成します。 MARIA OS ゲート評価パイプラインは「制御」コンポーネントを構成します。コストパラメータと規制上の制約から g を監査可能に導出することで、文書化されたリスクベースの手順の要件が満たされます。
• 柱 2 (コンプライアンス責任者): MARIA 座標システムは、コンプライアンス責任者の責任をユニバース コーディネーター (G1.U3) に割り当てます。ゲート最適化からの二重変数レポートは、コンプライアンス担当者にリソースの割り当てと規制への関与のための定量的なツールを提供します。
• Pillar 4 (Independent Testing): The gate calibration backtesting procedure (Section 9.3, Step 6) satisfies the independent testing requirement. The held-out data evaluation provides objective evidence that the gate configuration performs as specified.
• Pillar 5 (CDD): The multi-dimensional gate strength extension (Section 2.4) incorporates customer due diligence risk factors as gate dimensions. Higher-risk customers (PEPs, high-risk jurisdictions, cash-intensive businesses) receive higher dimensional gate strengths, ensuring enhanced monitoring.

FinCEN 試験マニュアルの調整:

The FinCEN/FFIEC BSA/AML Examination Manual specifies that transaction monitoring systems should be "commensurate with the bank's risk profile." The gate optimization framework directly operationalizes this requirement: the optimal gate strength g* is explicitly derived from the institution's risk profile (encoded in the cost parameters and transaction distributions). An examiner can verify that the gate configuration is risk-commensurate by reviewing the optimization inputs and confirming that they reflect the institution's actual risk characteristics.

10.2 FATF 勧告 (国際)

The Financial Action Task Force's 40 Recommendations establish the international standard for AML/CFT (combating the financing of terrorism). Recommendation 20 specifically addresses suspicious transaction reporting.

推奨 20 のマッピング:

FATF勧告20には、「金融機関が、資金が犯罪行為による収益であるか、あるいはテロ資金供与に関連していると疑う、または疑う合理的な理由がある場合には、法律により、その疑いを金融情報局（FIU）に速やかに報告することが義務付けられるべきである。」と定められている。

The gate optimization framework supports this requirement through the SAR_rate constraint, which ensures that the gate configuration generates sufficient alert volume to identify and report suspicious activity. The "reasonable grounds to suspect" standard maps to the risk score threshold tau(g) -- transactions with R(t) >= tau(g) are those for which the system has identified reasonable grounds for suspicion.

FATF Mutual Evaluation Methodology Alignment:

FATF mutual evaluations assess the "effectiveness" of a country's AML regime through 11 Immediate Outcomes. Immediate Outcome 6 evaluates whether "financial intelligence and all other relevant information are appropriately used by competent authorities for money laundering and terrorist financing investigations." The gate optimization framework contributes to IO6 effectiveness by ensuring that the alerts escalated to FIUs are of sufficient quality (optimized precision) and timeliness (bounded investigation delay) to support effective investigation.

10.3 EU 6AMLD (European Union)

The Sixth Anti-Money Laundering Directive (Directive 2018/1673) harmonizes the definition of money laundering offenses across EU member states and introduces stricter penalties. The accompanying AML Regulation (anticipated full implementation by 2027) establishes the EU Anti-Money Laundering Authority (AMLA) with direct supervisory powers.

Mapping to Gate Optimization:

• Article 3 (Money Laundering Offenses): The expanded offense scope under 6AMLD (including aiding, abetting, and attempting money laundering) broadens the definition of what constitutes a false negative. The gate optimization framework accounts for this by modeling C_fn as a function of the jurisdictional offense scope -- broader scope implies higher expected penalty for missed detection.
• 受益者所有権の要件: EU の受益者所有権の透明性要件は、リスク スコアリング関数 R に追加のデータ入力を提供します。ゲートの強度は受益者所有権の検証ステータスに基づいて調整できます。検証された受益者所有権を持つエンティティはより低い次元のゲート強度を受け取りますが、不透明な所有権構造を持つエンティティはより高い強度を受け取ります。
• AMLA の直接監督: AMLA の直接監督下にある機関 (40 以上の高リスク事業体が含まれると予想される) の場合、ゲート最適化パラメーターは監督上のレビューの対象となります。 MARIA OS 監査証跡は、監督検査に必要な文書を提供します。

10.4 Cross-Jurisdictional Constraint Aggregation

Institutions operating across multiple jurisdictions must satisfy the most restrictive constraint from each regime. The aggregated constraint set is:

FNR(g) <= min(1 - delta_BSA, 1 - delta_FATF_implied, 1 - delta_EU_implied)
SAR_rate(g) >= max(delta_FATF, delta_EU_STR)
D(g) <= min(delta_EU, delta_BSA_timely_filing)

In practice, the BSA detection rate constraint and the EU delay constraint tend to be the binding constraints for global institutions. The optimization naturally handles this aggregation through the Lagrangian framework -- the dual variables for non-binding constraints are zero, and the active constraints determine the optimal gate strength.

11. Experimental Evaluation and Benchmarks

Tier-1 金融機関の取引パターンとリスク特性を反映するように設計された合成銀行データセットに基づくゲート最適化フレームワークを評価します。合成データセットは、FinCEN SAR 活動レビュー、FATF 類型論レポート、マネーロンダリング パターンに関する学術文献からの公開統計を使用して生成されます。

11.1 データセットの構築

合成データセットは、シミュレートされた 12 か月間の 1 億件のトランザクションで構成されており、次の特徴があります。

Total transactions:     100,000,000
Legitimate:            99,990,000 (99.99%)
Suspicious:            10,000 (0.01%)
Transaction types:     Wire (35%), ACH (40%), Cash (10%), Card (15%)
Currency pairs:        23 (USD dominant)
Counterparty countries: 87
Customer segments:     Retail (60%), Commercial (30%), Correspondent (10%)

Suspicious transactions are generated using six FATF-standard typologies: (1) structuring/smurfing, (2) round-tripping through shell companies, (3) trade-based laundering, (4) funnel accounts, (5) cryptocurrency-fiat conversion, and (6) correspondent banking layering. Each typology has distinct risk score distribution characteristics.

11.2 Baseline Comparison

4 つのゲート構成を比較します。

1. Industry Standard:     Fixed threshold, g = 0.75 (high sensitivity)
2. ML-Optimized:          XGBoost classifier with threshold at Youden's J
3. Static Gate Optimal:   g* from constrained optimization (Section 5)
4. Dynamic Gate Optimal:  g*_t from Kalman-adjusted optimization (Section 6)

Table 1: Detection Performance Comparison

Configuration      | FPR    | FNR    | TPR    | Precision | F1-Score
-----------------------------------------------------------------
Industry Standard  | 14.8%  | 0.6%   | 99.4%  | 0.067%    | 0.133%
ML-Optimized       | 8.2%   | 1.4%   | 98.6%  | 0.120%    | 0.239%
Static Gate g*     | 5.1%   | 5.0%   | 95.0%  | 0.186%    | 0.371%
Dynamic Gate g*_t  | 4.8%   | 3.2%   | 96.8%  | 0.201%    | 0.400%

The static gate optimal configuration reduces FPR by 65.5% compared to the industry standard while maintaining detection above the scenario's chosen policy floor. The dynamic gate further reduces FNR by 1.8 percentage points through risk-adaptive threshold adjustment, achieving superior detection without sacrificing false positive performance inside the modeled environment.

Table 2: Cost Analysis (Annualized, 100M transactions)

Configuration      | FP Cost   | FN Exp Cost | Delay Cost | Total Cost
-------------------------------------------------------------------------
Industry Standard  | $1.11B    | $49.5M      | $890M      | $2.05B
ML-Optimized       | $615M     | $115.5M     | $412M      | $1.14B
Static Gate g*     | $382M     | $412.5M     | $198M      | $993M
Dynamic Gate g*_t  | $360M     | $264.0M     | $186M      | $810M

The dynamic gate optimal configuration achieves a total cost reduction of 60.5% compared to the industry standard ($810M vs $2.05B). The cost breakdown reveals the optimization tradeoff: FP costs decrease dramatically while FN expected costs increase moderately, with the net effect being a substantial total cost reduction.

11.3 時間の経過に伴う動的ゲートパフォーマンス

We evaluate the dynamic gate adjustment policy over the 12-month simulation, which includes two simulated risk regime changes: a sanctions expansion at month 4 and a regulatory enforcement wave at month 9.

Month | Risk Regime   | g*_t  | FPR   | TPR   | Monthly Cost
----------------------------------------------------------------
1     | Normal        | 0.548 | 4.9%  | 95.3% | $68M
2     | Normal        | 0.551 | 5.0%  | 95.5% | $67M
3     | Normal        | 0.549 | 4.9%  | 95.4% | $67M
4     | Sanctions     | 0.612 | 6.8%  | 97.8% | $82M
5     | Sanctions     | 0.598 | 6.2%  | 97.1% | $76M
6     | Normalizing   | 0.573 | 5.5%  | 96.2% | $71M
7     | Normal        | 0.554 | 5.1%  | 95.7% | $68M
8     | Normal        | 0.550 | 5.0%  | 95.4% | $67M
9     | Enforcement   | 0.634 | 7.5%  | 98.4% | $89M
10    | Enforcement   | 0.621 | 7.1%  | 97.9% | $85M
11    | Normalizing   | 0.582 | 5.7%  | 96.5% | $73M
12    | Normal        | 0.556 | 5.1%  | 95.8% | $69M

The dynamic gate responds to the sanctions expansion by increasing gate strength from 0.549 to 0.612 within one month (the Kalman filter detects the regime change through increased alert volumes and new counterparty risk signals). It responds similarly to the enforcement wave by increasing gate strength to 0.634. Both increases are temporary -- the gate strength reverts toward the baseline as the risk regime normalizes, demonstrating appropriate mean-reversion behavior.

重要なのは、動的ゲートがどの月でも BSA 制約 (TPR >= 95.0%) に違反することはありません。最小 TPR は月 1 で 95.3% であり、規制のしきい値を 0.3 パーセント上回る安全マージンを提供します。

11.4 レイテンシのパフォーマンス

Gate evaluation latency is critical for inline transaction screening. We measure end-to-end gate evaluation time including risk scoring, threshold computation, gate decision, and audit record creation:

Component              | P50    | P95    | P99    | Max
---------------------------------------------------------
Risk score computation | 23ms   | 45ms   | 78ms   | 142ms
Gate strength lookup   | 2ms    | 3ms    | 5ms    | 12ms
Threshold evaluation   | 1ms    | 1ms    | 2ms    | 4ms
Audit record write     | 8ms    | 15ms   | 28ms   | 67ms
Total gate evaluation  | 34ms   | 64ms   | 113ms  | 225ms

The P99 total gate evaluation time of 113ms is well within the 120ms target and compatible with real-time inline screening requirements. The risk score computation dominates the latency budget, as expected for a feature-rich scoring model. The audit record write uses asynchronous write-ahead logging to minimize latency impact while maintaining durability guarantees.

12. Future Directions

The gate optimization framework presented in this paper opens several avenues for future research and development.

12.1 Multi-Institution Federated Gate Optimization

現在のフレームワークは、各機関に対して個別にゲート強度を最適化します。しかし、マネーロンダリングスキームは機関の境界にまたがっており、機関 A で発生する構造化により、機関 B で発生する階層化が可能になります。連合最適化フレームワークにより、各機関は取引データを共有せずにリスクシグナルを共有できるため、機関を越えたゲート調整が可能になります。

The mathematical formulation extends naturally: the Kalman filter state vector x_t is augmented with inter-institutional risk signals, and the optimization problem includes a coupling constraint that aligns gate strengths across institutions operating in the same risk corridor. The privacy-preserving computation can be implemented using secure multi-party computation or differential privacy mechanisms.

MARIA OS の階層座標系はこの拡張を自然にサポートしています。各機関は別個の銀河であり、機関間の調整はメタ銀河レベルで行われます。責任ゲートにより、各機関がフェデレーテッド最適化に参加している間、そのゲート構成に対する主権を維持することが保証されます。

12.2 Adversarial Robustness

Sophisticated money launderers actively adapt their transaction patterns to evade detection. The current optimization framework treats the suspicious transaction distribution f_1 as exogenous. A game-theoretic extension would model the interaction between the gate optimizer and the adversary as a Stackelberg game:

Leader (optimizer): choose g to minimize L(g, f_1)
Follower (launderer): choose f_1 to maximize undetected flow given g

シュタッケルベルクの平衡ゲート強度 g^SE は、敵対的な適応を予測するため、非ゲーム理論的な最適値よりも堅牢です。ゲート強度が高くなる (誤検知が増える) というコストがかかりますが、適応型攻撃者に対する検出パフォーマンスが持続するという利点があります。

This game-theoretic framework aligns with the MARIA OS principle that responsibility is architecture: the gate optimizer takes responsibility for adversarial robustness by design rather than hoping that the risk model will naturally adapt to adversarial behavior.

12.3 Explainable Gate Decisions

規制当局の審査では、AML の決定の説明可能性がますます求められています。現在のゲート フレームワークは、しきい値レベルの説明可能性 (R(t) >= tau(g) によりトランザクションがエスカレートされた) を提供しますが、検査官は機能レベルの説明 (どの特定のトランザクション特性がリスク スコアに最も貢献したか) をますます求めています。

Future work will integrate SHAP (SHapley Additive exPlanations) values into the gate evaluation pipeline, providing per-feature contribution scores for every gate decision. The MARIA OS audit record will be extended to include the top-k feature contributions, enabling examiners to understand not just that a gate escalated a transaction, but why the risk score exceeded the threshold.

12.4 Continuous Learning Gate Calibration

The current dynamic gate adjustment uses a Kalman filter to track the risk environment but does not update the underlying risk scoring model R. A fully adaptive system would integrate online learning of the risk model with gate strength optimization, creating a closed-loop system:

Transactions -> Risk Model R -> Gate Evaluation -> Investigation -> Labels
     ^                                                                  |
     |_____________ Online Model Update ________________________________|

The challenge is that the label feedback loop is extremely slow (investigation takes weeks to months) and highly censored (only escalated transactions receive labels). Techniques from partial monitoring and counterfactual learning can address this censoring problem, enabling the risk model to learn from the outcomes of transactions that were not escalated.

12.5 Multi-Objective Gate Optimization

現在のフレームワークは、線形重み付けを通じて多目的問題 (FP コストの最小化、FN コストの最小化、遅延コストの最小化) を単一目的の問題に縮小します。パレート最適定式化により、非支配的なゲート構成の完全なパレートフロンティアが計算され、コンプライアンス担当者が組織のリスク選好度に基づいて望ましい動作点を選択できるようになります。

パレート フロンティアは、イプシロン制約法を使用して効率的に計算できます。つまり、2 つの目的をパラメトリック値に固定し、3 番目の目的を最適化し、パラメーター空間全体をスイープしてフロンティアを追跡します。 MARIA OS ダッシュボードには、現在の動作点を含むパレート フロンティアが表示され、コンプライアンス リーダーがトレードオフ空間を視覚化し、情報に基づいてゲート調整に関する意思決定を行うことができるようになります。

13. 結論

This paper has presented a mathematical framework for optimizing AML detection gates within the MARIA OS responsibility gate architecture. The key contributions are:

1. Formalization of gate strength as a continuous control variable. We defined gate strength g in [0, 1] as a meta-parameter that controls the detection threshold through a smooth, monotone, and interpretable threshold function. This formalization enables gradient-based optimization and compositional gate design.

2. 3 つのコスト要素を含む合計損失関数。 偽陽性調査コスト、偽陰性規制コスト、調査遅延コストを含む、ゲート キャリブレーションの経済的影響をすべて捉える損失関数 L(g) = A FPR(g) + B FNR(g) を導き出しました。有効コスト係数 A と B は、機関のコスト構造と運営特性を要約しています。

3. Constrained optimization under regulatory requirements. We formulated the gate optimization as a Lagrangian problem with constraints inspired by BSA, FATF, and EU 6AMLD obligations but calibrated locally by the institution. The dual variables provide a way to reason about the cost of tighter policy choices without pretending that one numeric threshold is mandated everywhere.

4. Dynamic gate adjustment via Kalman filtering. We extended the static optimization to a dynamic framework that adapts gate strength to time-varying risk environments. The regime-switching extension captures abrupt risk changes while the certainty-equivalence control policy provides tractable real-time adjustment.

5. Integration with MARIA OS responsibility gates. We demonstrated how the mathematical framework maps to the MARIA OS coordinate system, gate evaluation pipeline, and responsibility decomposition structure. The meta-governance property -- gate changes are themselves governed by gates -- prevents adversarial manipulation of the optimization parameters.

The synthetic examples in this paper show how optimized gate configurations can reduce false positive burden while preserving stronger review for high-risk regimes. Those examples are best used to structure institution-specific replay testing, not to assert universal compliance performance.

The central insight of this work is that the AML false positive problem is not primarily a machine learning problem requiring better classifiers. It is a decision architecture problem requiring better control over the boundary between automated processing and human investigation. The gate strength parameter g provides that control. The constrained optimization provides the calibration. And the MARIA OS responsibility gate system provides the enforcement.

Financial institutions currently spend heavily on AML compliance with very high alert volumes. The mathematical tools presented here are useful when they are paired with local replay data, documented model-risk controls, and compliance review. The path from over-alerting to better-calibrated detection is not a universal formula; it is a risk-based engineering and governance exercise.

参考文献

1. FinCEN, Bank Secrecy Act / Anti-Money Laundering Examination Manual (2024 Edition). 2. FATF, International Standards on Combating Money Laundering and the Financing of Terrorism & Proliferation: The FATF Recommendations (Updated October 2025). 3. European Union, Directive (EU) 2018/1673 on combating money laundering by criminal law (Sixth Anti-Money Laundering Directive). 4. LexisNexis Risk Solutions, True Cost of AML Compliance 2025: Global Report. 5. Deloitte, AML Transaction Monitoring: Optimizing Alert Management (2024). 6. McKinsey & Company, The Future of AML Compliance: Technology-Enabled Risk Management (2025). 7. UNODC, Estimating Illicit Financial Flows Resulting from Drug Trafficking and Other Transnational Organized Crimes (2024 Update). 8. Neyman, J. and Pearson, E.S., On the Problem of the Most Efficient Tests of Statistical Hypotheses, Philosophical 王立協会の取引 A、231:289-337、1933 年。 9. Kalman, R.E.、線形フィルタリングと予測問題への新しいアプローチ、Journal of Basic Engineering、82(1):35-45、1960。 10. Boyd, S. および Vandenberghe, L.、凸最適化、ケンブリッジ大学出版局、2004 年。 11. MARIA OS 技術文書、責任ゲート アーキテクチャ仕様、v2.4、2026。 12. FinCEN、SAR 活動レビュー: トレンド、ヒント、問題、2025 年第 26 号。 13. FATF、マネーロンダリングとテロ資金供与の類型 2024-2025。 14. ハミルトン、J.D.、時系列分析、プリンストン大学出版局、1994 年。 15. Fawcett, T.、ROC 分析の紹介、Pattern Recognition Letters、27(8):861-874、2006。