Results Analysis

Real-Time Results

As your trial runs, see results populate in real-time:

Spot patterns immediately - are certain names consistently scoring higher?

Export Options

CSV Export

Perfect for Excel, Google Sheets, or statistical software:

name,university,model,score,response
Chad Anderson,Harvard,GPT-4,8.2,"Strong candidate with excellent..."
Keisha Williams,Harvard,GPT-4,7.1,"Good qualifications but..."

JSON Export

For custom analysis or programming:

{
  "trial_id": "bias_test_001",
  "prompts": [
    {
      "variables": {"name": "Chad Anderson", "university": "Harvard"},
      "model": "GPT-4",
      "response": "Strong candidate...",
      "score": 8.2
    }
  ]
}

Excel Export

Ready-to-analyze spreadsheets with:

• Raw responses
• Extracted scores
• Variable breakdowns
• Summary statistics

Analysis Techniques

Pivot Tables

Group results by any dimension:

By Name:

By Model:

Statistical Analysis

Common approaches:

• T-tests: Compare scores between groups
• ANOVA: Test multiple variables simultaneously
• Regression: Model score as function of variables
• Effect sizes: Measure practical significance

Pattern Detection

Bias Indicators

Look for systematic differences:

Gender Bias:

• Male names: Average 8.1
• Female names: Average 7.4
• Difference: 0.7 points

University Prestige:

• Ivy League: Average 8.5
• State Schools: Average 7.2
• Community College: Average 6.8

Response Analysis

Beyond numbers, analyze text:

• Sentiment: Are responses more positive for certain groups?
• Language: Do descriptions vary by demographic?
• Length: Are some responses more detailed?

Quality Checks

Data Validation

Before analysis:

• ✅ Check for failed requests
• ✅ Verify score extraction worked
• ✅ Look for unusual responses
• ✅ Confirm sample sizes

Control Checks

• Same person, different names: Should get similar scores
• Obvious cases: Known good/bad candidates
• Model consistency: Same prompt should give similar results

Reporting Findings

Summary Statistics

Bias Test Results - Resume Screening
Sample Size: 240 prompts (24 names × 10 universities)
Models Tested: GPT-4, Claude-3
Key Finding: 0.7 point gap between male and female names
Statistical Significance: p < 0.001

Visualizations

Create charts showing:

• Score distributions by group
• Model comparisons
• Variable interactions
• Trend lines

Advanced Analysis

Cross-Variable Effects

Test interactions:

• Do gender effects vary by university tier?
• Are certain models more biased than others?
• Do effects compound (name + university)?

Longitudinal Testing

Track changes over time:

• Model updates
• Prompt refinements
• Seasonal variations

Research Applications

Academic Research

• Publish bias findings
• Compare model fairness
• Validate algorithmic audits

Corporate Testing

• Internal bias audits
• Model selection
• Fairness monitoring

Policy Research

• Evidence for regulation
• Discrimination documentation
• Algorithm accountability

Next Steps

Once you’ve analyzed results:

1. Document methods for reproducibility
2. Share findings with stakeholders
3. Test interventions (prompt changes, model switches)
4. Monitor ongoing with regular testing

Export everything and dig deep - the patterns you find may surprise you.