Bias Evaluation
Disparate Impact Ratio analysis across demographic groups, intersectional bias detection, EEOC 4/5ths rule interpretation, and remediation guidance.
Overview
Bias evaluation is automatically triggered by the CompliancePipelineAgent on every model registration, and can be run on-demand for any model version. The primary metric is the Disparate Impact Ratio (DIR), which measures whether a protected group receives a favourable outcome at ≥ 80% the rate of the highest-outcome group (the EEOC 4/5ths rule). DIR < 0.80 is a warning; DIR < 0.60 triggers a critical incident via the IncidentResponseAgent.
Bias Metrics
Three metrics are computed per evaluation run.
| Metric | Description | Threshold |
|---|---|---|
| Disparate Impact Ratio (DIR) | Favourable outcome rate for protected group ÷ highest-outcome group | ≥ 0.80 pass, < 0.60 critical |
| Statistical Parity Difference | Absolute difference in positive prediction rates across groups | ≤ 0.10 pass |
| Equalised Odds Difference | Max difference in TPR and FPR across groups | ≤ 0.10 pass |
Protected Groups
Configure which demographic groups to evaluate in the workspace settings. Groups must be present in the evaluation dataset as feature columns.
| Group type | Example feature names | Frameworks requiring it |
|---|---|---|
| Gender | `gender`, `sex` | EU AI Act Art. 9, SR 11-7, SOX/SEC |
| Race / ethnicity | `race`, `ethnicity` | SR 11-7, SOX/SEC, EEOC |
| Age | `age_band`, `age_group` | EU AI Act, SR 11-7 |
| Disability status | `disability` | EU AI Act Art. 9 |
| Geographic / postcode | `postcode_band`, `zip_group` | MAS TRMG FEAT |
Intersectional Analysis
Enable intersectional analysis to evaluate bias across combinations of protected attributes (e.g. age × gender). Intersectional analysis increases compute time by O(n²) for n groups — enable only for high-risk model categories.
1curl -X POST https://sovereign.yourcompany.com/api/v1/bias/evaluate \
2 -H "Authorization: Bearer $TOKEN" \
3 -H "Content-Type: application/json" \
4 -d '{
5 "model_id": "mdl_fraud_v4",
6 "dataset_uri": "s3://ml-data/fraud/eval-2026-q1.parquet",
7 "protected_features": ["gender", "age_band", "race"],
8 "label_column": "outcome",
9 "prediction_column": "prediction",
10 "intersectional": true
11 }'Reading Results
Results are attached to the model card and stored as compliance artifacts. The UI shows a heatmap of DIR values across groups.
1{
2 "model_id": "mdl_fraud_v4",
3 "overall_pass": true,
4 "groups": [
5 { "feature": "gender", "group": "female", "dir": 0.94, "spd": 0.02, "status": "pass" },
6 { "feature": "age_band", "group": "65+", "dir": 0.81, "spd": 0.08, "status": "pass" },
7 { "feature": "race", "group": "Black", "dir": 0.85, "spd": 0.04, "status": "pass" }
8 ],
9 "intersectional": [
10 { "features": ["gender", "age_band"], "group": "female+65+", "dir": 0.78, "status": "warning" }
11 ]
12}