Algorithm Catalog — Fraud, AI, and ML (Phase 0→1, with Phase 2 Roadmap)

Document Classification: Internal — CHLOM Confidential Owner: CrownThrive, LLC Last Updated: 2025-08-08

Overview

This catalog contains the fraud detection, AI, and machine learning algorithms planned and deployed for CHLOM’s Compliance-as-a-Service (CaaS) and Tokenized Licensing-as-a-Service (TLaaS) framework. It includes:

• Core algorithms and models with their purposes, inputs, outputs, pseudocode, KPIs, and performance targets.
• Security and privacy considerations.
• Phase 2 roadmap for scaling, improving accuracy, and integrating cryptographic verifications.
• Pre-filled Proprietary Algorithm Documentation for key codenames.
• Starter code stubs in Rust and Python (gRPC) for immediate prototyping.
• Implementation readiness notes so developers can build without guesswork.

Algorithm Inventory & Pre-Filled Docs

1. AegisScore-LR (Logistic Regression Compliance Scoring)

• Purpose: Assign real-time compliance risk scores to entities.
• Inputs: Entity type, identifiers, transaction history, sanctions/PEP results, ZK proofs.
• Outputs: Risk score (0–1), confidence interval.
• Core Logic: Logistic regression over normalized features; threshold-based classification.
• KPIs: Accuracy ≥ 92%, ROC-AUC ≥ 0.95.
• Security: Remove PII from logs, encrypt feature vectors.
• Endpoints:
• Testing: Cross-validation on multi-jurisdiction datasets.

2. AegisScore-GBT (Gradient Boosted Trees Variant)

• Purpose: Higher complexity compliance scoring for nuanced cases.
• Core Logic: XGBoost with tuned hyperparameters.
• KPIs: ROC-AUC ≥ 0.97, inference latency ≤ 150ms.
• Security: Feature encryption at rest, API auth with mTLS.

3. GBA-Prop (Graph-Based Anomaly Propagation)

• Purpose: Detect fraud rings and hidden relationships.
• Logic: Label propagation + temporal anomaly thresholds.
• KPIs: Precision ≥ 90%, recall ≥ 85%.
• Security: Encrypted Neo4j store, signed query access.

4. SanctionsMatcher-Hybrid

• Purpose: High-recall sanctions/PEP list matching.
• Logic: BERT NER → TF-IDF → cosine similarity.
• KPIs: Recall ≥ 98%.
• Security: No unencrypted storage of raw identifiers.

5. IsoForest-AD (Isolation Forest Anomaly Detection)

• Purpose: Unsupervised detection of outlier transactions.
• KPIs: FPR ≤ 5%.
• Security: Isolation of feature processing from public endpoints.

6. PR-Fraud (PageRank Fraud Influence Scoring)

• Purpose: Rank entities by fraud influence.
• Logic: Modified PageRank weighted by fraud signals.
• KPIs: Top-10 bad actor coverage ≥ 90%.

Starter Code Stubs

Python (gRPC Server)

# Compliance gRPC Server Stub
class ComplianceService(...):
    def ScoreEntity(self, request, context):
        # Load model, preprocess, score
        return ScoreResponse(score=0.85, confidence=0.92)

Rust (gRPC Client)

let mut client = ComplianceClient::connect("http://[::1]:50051").await?;
let response = client.score_entity(...).await?;

Shared .proto files define messages for AegisScore-LR, GBA-Prop, SanctionsMatcher-Hybrid, IsoForest-AD, and PR-Fraud endpoints.

Developer Implementation Readiness

• Pre-filled Algorithm Docs for all above codenames include:
• Purpose & scope
• Inputs/outputs (schemas)
• Pseudocode & diagrams
• KPIs & performance targets
• Security considerations
• API interface details
• Testing & validation plans
• Maintenance/versioning strategy
• Developers can build and deploy immediately without ambiguity.

Phase 2 Roadmap

1. Federated Learning for cross-border compliance.
2. Graph Neural Networks for deeper fraud detection.
3. Zero-Knowledge Model Verification.
4. Real-Time Drift Detection Pipelines.
5. Quantum-Resistant Signatures for outputs.

Next Step: Integrate into Tier 3 Secure Onboarding Bundle with restricted access controls.