AI governance · Australian population data
Same applicant, different sex. Does your AI’s answer change?
Find out in ten minutes. No code, no account.
pip install verosynthea-validator for the local version
Sample result
+9.2 percentage points
average change when only sex changes
Gap peaks: high income · manager
What the test does
- Measures your model's fairness for protected demographics
- Tests it against realistic Australian profiles
- Flags where it treats people differently
What you need
- Your scoring model (.pkl)The trained file that turns applicant details into a score or a decision.
- A one-page data dictionary (template)
- Ten minutes, no code
What you get
- Measures to assess how fair your model is
- A severity rating, and where the gaps peak
- A record for your model documentation
Measures direct sensitivity only; effects through correlated features are not visible here. Deliberate use of a protected attribute is documented, not judged.
Full scope and limits
This test measures direct sensitivity to protected attributes; discrimination that enters through correlated features such as occupation or suburb is not visible to it. And if your model uses a protected attribute deliberately, for example sex in insurance pricing where actuarial justification applies, the validator does not judge that choice. It documents the size of the effect so that you can.
- 1Model→
- 2Dictionary→
- 3Mapping→
- 4Buckets→
- 5Results
Upload your model
Your model file must:
- Have a
.predict()method - Take demographic features as input (age, sex, country of birth, plus others)
- Return a prediction (for example a classifier output)
Most scikit-learn / XGBoost / LightGBM models work out of the box.
New reports, occasionally.
One email when we publish new Census analysis. No other mail.