Unlocking Modern Fraud Defense: An Introduction to the Fraud Language Model (FLM)

Balancing Speed, Transparency, and Security in Real-Time Fraud Prevention

As financial fraud evolves to exploit real-time payment systems, banks and fintechs face the difficult task of stopping bad actors without harming legitimate customers. Legacy defenses fall short: rigid rules miss subtle patterns, while black-box AI creates opacity that regulators and operations teams cannot trust. The Fraud Language Model (FLM), developed by Loci, introduces a new model class purpose-built for fraud defense. FLM treats rules as models—scored, weighted, and explainable—delivering the intelligence of machine learning with the clarity of human-readable logic.

This article explains FLM in clear terms. What it is, how it works, its benefits, and its trade-offs—so business leaders and fraud teams can understand where it fits in a modern defense strategy.

FLM Explained: A Clear Framework for Complex Decisions

Imagine fraud detection as a conversation between humans and machines. Traditional rules are strict checklists: "If a transaction is over $10,000 from a new device, block it." While easy to understand, they are rigid and often miss subtle fraud patterns or generate excessive false alarms. At the other extreme, machine learning (ML) models act as opaque "black boxes." They excel at finding hidden correlations in data but often cannot explain their reasoning—a significant problem in regulated industries where transparency is mandatory.

FLM acts as a translator in this conversation. It is a specialized language for describing fraud patterns in a structured format that is both human-readable and machine-executable. At its core, FLM treats rules as flexible models, not binary switches. Its key building blocks include:

Evaluations: The specific risk signals being checked, such as transaction velocity in a short period or suspicious similarity between payment amounts.
Weights: The importance assigned to each signal by an analyst. A strong indicator (like a sudden spike in activity) receives a high weight, while a minor one gets less.
Thresholds: The minimum score required to trigger an action. This adds nuance, ensuring that not every minor red flag triggers an alarm.
Scores: A nuanced risk score (e.g., from 0 to 1) calculated from the weighted evidence, allowing for shades of gray instead of a simple pass/fail.
Actions: The specific outcome when a rule is triggered, such as blocking a transaction, flagging it for review, or escalating it to a specific team.

Because FLM rules are written in a simple data format like JSON, they are easy to version, audit, and replay. This allows operators to trace the exact logic behind any decision, which is essential for satisfying regulators under frameworks like the EU AI Act.

Unlike heavy ML systems that demand expensive hardware and specialized data scientists, FLM delivers decisions in milliseconds on standard servers. Its deterministic nature (the same input always produces the same output) builds trust, while still allowing for AI-driven enhancements like rule suggestions from tools that analyze historical data.

How FLM Works: A Real-World Example

Let's examine a common fraud tactic called "structuring" or "smurfing," where criminals split a large sum into smaller transfers to evade detection. A traditional rule might say: "Flag a customer if they make five or more transfers in 30 minutes totaling over ₦750,000."

In FLM, this becomes a more nuanced, structured model:

Evaluation 1: Count transfers in 30 minutes (Weight: 4, as high velocity is a strong indicator).
Evaluation 2: Check if transfer amounts are unusually similar (Weight: 4, indicating suspicious uniformity).
Evaluation 3: Sum the total amount transferred (Weight: 3, as the total is important but not definitive on its own).
Threshold: 0.75 (meaning at least 75% of the weighted evidence must be present).

Now, suppose a customer makes several quick, similar payments totaling ₦680,000. The velocity and amount similarity evaluations pass, but the total amount does not. The risk score is calculated as the weight of the passed evaluations divided by the total possible weight: (4 + 4) / (4 + 4 + 3) = 8 / 11 ≈ 0.73.

Since the 0.73 score is below the 0.75 threshold, the rule doesn't trigger a block. This prevents a legitimate customer from being disrupted. However, the system logs a "near miss," providing valuable data for future monitoring. For business leaders, the benefits are threefold: fewer false positives improve the customer experience, the threshold is configurable, and the detailed audit trail simplifies compliance.

Key Benefits for Financial Services

FLM excels in environments where speed, clarity, and cost are paramount.

Real-Time Protection: In instant payment systems like FedNow or NIP, sub-second decisions are non-negotiable. Loci evaluates complex FLM rules in under a second.
Reduced False Positives: Weighted scoring catches subtle patterns without overreacting, potentially cutting unnecessary transaction blocks by 20-30%.
Explainability and Compliance: In an era of increasing regulatory scrutiny, FLM’s transparency is a game-changer. Auditors can replay decisions, and customers can receive clear reasons for holds.
Ease of Use: Tools like Loci Studio empower fraud analysts to create and deploy models quickly, while features like Autographer suggest new rules based on data trends, keeping defenses adaptive.
Scalability: Encoded directly in JSON, FLM rules are machine-readable and system-agnostic, ensuring smooth adoption whether in a startup stack or enterprise core.

Real-World Use Cases

FLM is a practical tool for addressing everyday fraud challenges:

Instant Payment Fraud: Detects rapid "fan-out" patterns (money spreading to multiple accounts) and other velocity attacks without adding latency to legitimate transfers.
Account Takeover Detection: Combines multiple weak signals—such as a new device, an unusual location, and atypical login time—into a single, strong indicator of compromise.
Anti-Money Laundering (AML): Automates the detection and documentation of structuring and smurfing attempts, generating clear audit trails for regulatory reporting.
Democratizing Advanced Defense: Allows smaller banks or teams with limited resources to deploy sophisticated, weighted rules in minutes without specialized coding skills.

Limitations and Considerations: A Balanced View

FLM is a powerful tool, but not a universal solution. Its design ensures consistency and auditability, yet it does not continuously self-learn like machine learning. Updates come through fraud analysts or data-driven assistive tools such as Autographer, balancing human oversight with AI discovery.

Governance remains critical. Without clear oversight, “rule sprawl” can emerge, where overlapping or redundant rules reduce efficiency. However, Loci mitigates onboarding and authoring challenges with its natural-language-to-model pipeline and Studio interface, allowing analysts to describe rules in plain English that compile directly into FLM. This reduces setup friction compared to legacy systems.

FLM is not a middle ground, but a new model class. Legacy rules are rigid. ML is adaptive but opaque. FLM delivers transparency, precision, and regulatory trust, while still evolving through analyst input and AI-assisted authoring. For institutions prioritizing explainability, speed, and cost-efficiency in real-time payments, FLM stands as a complete detection model—ready for production without the black-box compromises of machine learning.

The Future of FLM: Toward Explainable Intelligence

As fraudsters arm themselves with AI, the defense landscape is shifting toward "explainable intelligence"—a blend of human oversight and technological precision. With features like shadow testing (running rules silently to measure their effectiveness) and AI-driven rule suggestions, the FLM framework is poised to become a core component of the modern fintech stack.

For any organization in financial services, the path forward requires balancing innovation with trust. By weighing FLM's strengths in transparency and speed against your unique risk appetite and adaptability needs, you can make an informed decision. In an industry where every second and every decision counts, FLM offers a balanced approach that empowers teams without the mystery of a black box.

This article is based on publicly available information from Loci's documentation on FLM. For implementation advice, consult fraud experts or Loci directly. Also feel free to take the Fraud Defense Self-assessment Quiz