Outcome balance
Selection rates and true-positive rates at the active threshold
Group AGroup BEqual bars do not imply equal error rates
Hugdrigo Portfolio / Model Governance / FairnessScope
Decision systems · fairness diagnostics
Measure the disparity.
Understand the trade-off.
Explore transparent classification metrics across two synthetic groups. Inspect threshold choices, errors and group outcomes without claiming that any single metric proves fairness.
● SIMULATED DATA
Threshold 50%
Diagnostic summary
Absolute gaps between groups
A disparity is a prompt to investigate, not proof of unlawful discrimination. Base rates, labels and measurement quality matter.
Group confusion matrices
Errors and correct predictions by cohort
Threshold sensitivity
Compare selection disparity as the decision cutoff moves
Group A selectionGroup B selection
Scenario testing
Threshold laboratory
Set a decision cutoff, assign relative error costs and compare policy choices on the same synthetic population.
Decision controls
Changes update every metric immediately
This tool compares threshold-based decisions. It does not establish legal compliance, recommend deployment, or evaluate causes of group disparities.
Cost vs disparity
Changing the threshold changes both error costs and selection gap
Normalized error costSelection gap, pp
Threshold comparison
Compare preconfigured options and current custom setting
| Policy | Threshold | Selection gap | TPR gap | Overall accuracy | Weighted error cost |
|---|
Cohort diagnostics
Group explorer
Compare observed performance and error distributions. The fictional groups are generic operational cohorts, not demographics.
Classification outcomes
Counts and rates at the active threshold
| Group | N | Base rate | Selection | TPR | FPR | Precision | Accuracy |
|---|
Observed outcome composition
TP, FP, FN and TN counts
Notes for interpretation
Questions worth asking before acting
1
How were labels collected?Historical outcomes can contain missingness, noise and pre-existing bias.
2
Do base rates differ?Conflicting group-fairness criteria may not all be jointly achievable.
3
Is the sample representative?Small groups and sampling bias can distort apparent gaps.
4
Are effects measured after deployment?Real governance requires stakeholders, monitoring and impact review.
Data stewardship
Synthetic record library
Create, edit, archive or restore fictional classifier observations. Your modifications remain in this browser unless you export a JSON backup.
Record registry
0 entries
| ID | Group | Model score | Observed label | Decision | Status | Actions |
|---|
Transparency & limitations
Methodology
Audit the assumptions behind this educational classifier-comparison application.
Metric definitions
Prediction: positive when score ≥ threshold.
Selection rate: predicted positives / group total.
True positive rate (TPR): TP / (TP + FN).
False positive rate (FPR): FP / (FP + TN).
Precision: TP / (TP + FP).
Accuracy: (TP + TN) / group total.
Demographic parity gap: absolute difference in group selection rates.
Equal opportunity gap: absolute difference in group TPR.
Equalized odds diagnostic: max of absolute TPR and FPR differences.
Error cost: FP × FP cost + FN × FN cost (illustrative units).
Safeguards and limitations
✓
Synthetic data onlyGenerated with a deterministic pseudorandom function. No sensitive attributes or real persons.
✓
No fairness certificationDifferent definitions can conflict. Metrics depend on label quality, setting and social context.
✓
Not a trained classifierOnly a simulated score and binary label dataset is analyzed. There is no online inference.
✓
Local-firstNo network calls, cookies, ads, external fonts, telemetry or servers. JSON backup is user-controlled.
✓
Defensive inputsValidates bounded JSON datasets; escapes HTML and neutralizes common spreadsheet formulas in CSV exports.
Portfolio project #112
Built as a reproducible demonstration of transparent AI evaluation for the Hugdrigo portfolio. Threshold experiments are descriptive, not policy advice. The datasets are fictional and should never be used as evidence about real groups or individuals.