Portal Walkthrough: Control 2.11 - Bias Testing and Fairness Assessment
Last Updated: January 2026 Portal: Custom testing environment, Power BI for analysis Estimated Time: 8-16 hours for initial assessment
Prerequisites
- Protected class definitions documented per ECOA
- Test dataset with demographic distribution
- Fairness metrics defined
- Data science team engagement
Step-by-Step Configuration
Step 1: Define Protected Classes
Document protected classes per ECOA and state law:
| Protected Class | Attribute | Data Source |
|---|---|---|
| Race | Demographics | Customer profile |
| Color | Demographics | Customer profile |
| Religion | Demographics | Customer profile |
| National Origin | Demographics | Customer profile |
| Sex | Demographics | Customer profile |
| Age | Date of birth | Customer profile |
| Marital Status | Demographics | Customer profile |
| Public Assistance | Income source | Application data |
| Good Faith Exercise of Consumer Credit Protection Act Rights | Rights exercise | Application data |
Step 2: Create Test Dataset
- Build representative test dataset:
- Minimum 1,000 test cases per protected class
- Balance across demographic groups
- Use synthetic data (not production customer data)
- Create standard prompt templates for testing
- Document dataset methodology
Step 3: Establish Fairness Metrics
Define metrics to measure:
| Metric | Definition | Threshold |
|---|---|---|
| Demographic Parity | Equal positive outcome rate across groups | ±5% |
| Equalized Odds | Equal true positive/false positive rates | ±5% |
| Calibration | Predicted probability matches actual outcomes | ±10% |
Step 4: Execute Bias Testing
- Run each test case through agent
- Capture agent response
- Classify response as positive/negative outcome
- Calculate metrics by demographic group
Step 5: Document and Remediate
- Generate bias testing report with statistics
- Identify significant disparities
- Create remediation plan for bias issues
- Re-test after remediation
Configuration by Governance Level
| Setting | Baseline (Zone 1) | Recommended (Zone 2) | Regulated (Zone 3) |
|---|---|---|---|
| Testing Frequency | Annual | Pre-deployment | Pre-deployment + Quarterly |
| Test Dataset Size | 500/group | 1,000/group | 2,000/group |
| Metrics | Demographic parity | + Equalized odds | Comprehensive |
| Documentation | Summary | Full report | Independent validation |
| Remediation SLA | 30 days | 14 days | 7 days critical |
FSI Example Configuration
Bias Testing: Investment Advisory Bot
Protected Classes:
- Race (5 categories)
- Sex (2 categories)
- Age (4 brackets: 18-35, 36-50, 51-65, 65+)
Test Dataset:
Total Size: 8,000 cases
Per Group: 1,000 minimum
Source: Synthetic generated
Fairness Metrics:
- Demographic Parity: Pass/Fail at ±5%
- Equalized Odds: Pass/Fail at ±5%
- Calibration: Pass/Fail at ±10%
Testing Schedule:
- Pre-deployment: Required
- Quarterly: Required for Zone 3
- After significant changes: Required
Validation
After completing these steps, verify:
- Protected classes documented per ECOA
- Test dataset created with representation
- Fairness metrics established
- Initial bias testing completed
- Results documented with remediation plan
Back to Control 2.11 | PowerShell Setup | Verification Testing | Troubleshooting