docs/DECISION_LOGIC.md

Decision Logic Documentation

Overview

FraudSimulator-AI implements a multi-stage decision intelligence system for insurance fraud detection. The system answers a single executive decision question:

"Should this insurance claim be investigated or allowed — and what evidence supports that decision?"

Decision Contract

Input

Structured claim data including:

• Claim metadata (ID, type, amount)
• Claimant history
• Policy information
• Document data
• Temporal patterns
• Entity relationships

Output

Binary decision with evidence:

{
  "decision": "investigate | allow",
  "fraud_score": 0.0-1.0,
  "risk_band": "low | medium | high",
  "evidence": ["list of fraud indicators"],
  "confidence": 0.0-1.0,
  "audit_id": "unique identifier",
  "timestamp": "ISO 8601 timestamp"
}

Decision Pipeline

Stage 1: Feature Engineering

Extract and normalize features from raw claim data:

• Amount features: Claim amount, deviation from average
• Frequency features: Claim count, time between claims
• Temporal features: Days since policy inception, claim timing
• Document features: Document completeness, consistency scores
• Entity features: Linked entities, relationship networks

Stage 2: Multi-Agent Analysis

Pattern Analysis Agent

Identifies fraud patterns:

• High Frequency: Claimant has submitted multiple claims in short period
• Amount Deviation: Claim amount significantly differs from historical average
• Early Claim: Claim filed shortly after policy inception (< 30 days)

Anomaly Detection Agent

Detects statistical anomalies:

• Document Anomalies: Missing or inconsistent documentation
• Entity Linkage: Connections to known suspicious entities
• Behavioral Anomalies: Unusual claim submission patterns

Risk Scoring Agent

Calculates weighted fraud risk score:

fraud_score = (pattern_score × 0.6) + (anomaly_score × 0.4)

where:
  pattern_score = (frequency × 0.4) + (amount_deviation × 0.3) + (temporal × 0.3)
  anomaly_score = (document × 0.4) + (entity × 0.4) + (behavioral × 0.2)

Stage 3: Decision Threshold

Apply decision threshold to fraud score:

• fraud_score ≥ 0.65: Recommend "investigate"
• fraud_score < 0.65: Recommend "allow"

Stage 4: Risk Banding

Classify risk level:

• High Risk: fraud_score ≥ 0.7
• Medium Risk: 0.4 ≤ fraud_score < 0.7
• Low Risk: fraud_score < 0.4

Stage 5: Explainability Generation

Build evidence list from activated indicators:

• List all indicators with score > 0.1
• Provide human-readable descriptions
• Include indicator weights
• Calculate decision confidence

Stage 6: Governance & Audit

Create audit trail:

• Generate unique audit ID
• Log timestamp (UTC)
• Record claim ID
• Store decision and evidence
• Track model version

Decision Confidence

Confidence is calculated based on indicator consistency:

variance = Σ(indicator_value - 0.5)² / n_indicators
confidence = 1.0 - (variance × 0.5)
confidence = max(confidence, 0.5)  // minimum 50% confidence

Higher confidence indicates:

• Indicators are aligned (all high or all low)
• Clear fraud pattern or clear legitimate pattern
• Less ambiguity in decision

Lower confidence indicates:

• Mixed signals from different indicators
• Borderline case requiring human review
• Potential for false positive/negative

Human-in-the-Loop Integration

The system is designed for human oversight:

1. High-confidence "investigate": Immediate escalation to fraud investigation team
2. Low-confidence "investigate": Flag for senior adjuster review
3. High-confidence "allow": Auto-approve with audit trail
4. Low-confidence "allow": Route to standard claims processing with monitoring

Model Versioning

Current version: 1.0.0

All decisions are tagged with model version for:

• Reproducibility
• A/B testing
• Regulatory compliance
• Drift detection

Regulatory Alignment

Decision logic complies with:

• IFRS 17: Insurance contract accounting standards
• AML Requirements: Anti-money laundering detection
• Explainability Standards: All decisions are explainable and auditable
• Bias Monitoring: Regular review of decision patterns across demographics

Performance Metrics

Target metrics:

• Precision: ≥ 75% (minimize false positives)
• Recall: ≥ 80% (catch majority of fraud)
• F1 Score: ≥ 0.77
• Decision Time: < 2 seconds per claim
• Explainability Coverage: 100% (all decisions explained)

Continuous Improvement

Decision logic is updated based on:

• Fraud investigation outcomes
• False positive/negative analysis
• Emerging fraud patterns
• Regulatory changes
• Stakeholder feedback