Assurance¶

An auditor's next questions after what AI do you run are why did it decide that, is it fair, and is it still behaving the way it did when you approved it. Assurance answers those three for one system, or for all of them, "in the language your auditor will accept".

Every result is computed from the predictions you logged, not from a template. Each tab leads with a plain-language verdict, has a How was this computed? note, and lets analysts and administrators Mark reviewed. With too little data, the tab says the result can't be computed yet rather than estimating one.

Use Choose a system (or All systems) at the top. The tabs are Explainability, Fairness and Drift.

Explainability¶

Explanations run your model

SHAP, LIME and DiCE explanations are computed by running the model itself, so WhiteBoxXAI needs the model file. Register the model through GitHub with storage_mode: full; scikit-learn-style pickle and joblib files are supported. For a model registered with metadata only, the explanation is recorded as failed, with a message saying no stored model artifact is available. Drift, fairness audits and reports don't need the model file.

Request an explanation¶

Choose a system, then a logged prediction to explain.

Pick a Method:

Method	Use it for
Model default	The method configured for this system
SHAP (kernel)	Any model; model-agnostic Shapley values
SHAP (tree models)	Tree ensembles such as XGBoost, LightGBM and random forests; exact and fast
LIME (tabular)	A local linear approximation around the prediction
DiCE (counterfactuals)	What would have to change for a different outcome?

Request the explanation. Results appear when the computation finishes.

Read the result¶

Contribution waterfall: how each feature moved the prediction away from the base value
Force plot: the features pushing the prediction up and down, side by side
Feature weights: the ranked contributions
What this explanation shows: a plain-language reading of the result

Counterfactuals (DiCE) answer the question an applicant or customer actually asks: what would have needed to be different? Each counterfactual is an alternative input, close to the original, that the model scores differently. From the SDK you can choose how many to generate (total_cfs, default 3), the target outcome (desired_class, default "opposite") and which features may change (features_to_vary, default "all"). Restrict features_to_vary to features a person could realistically change. Counterfactuals that vary age or protected attributes aren't useful recourse.

cf = client.explanations.generate(
    model_id,
    instance={"income": 52000, "utilisation": 0.31, "tenure_months": 18},
    method="dice",
    total_cfs=3,
    features_to_vary=["income", "utilisation"],
)

For how SHAP and LIME values are computed, and global feature importance, see Explainability Engine.

Fairness¶

Choose a Protected attribute (All, Gender, Race/Ethnicity, Age, Disability), then read the sub-tabs:

Sub-tab	Shows
Fairness Metrics	Demographic parity, equalized odds, disparate impact and the other metrics computed by the audit
Group Comparison	Outcomes and error rates per group, side by side
Audit Results	Each recorded audit, its severity, and whether it passed
Recommendations	The mitigations the audit recorded for this result

Where audit data comes from¶

A fairness audit needs true outcomes, predicted outcomes and group membership. WhiteBoxXAI doesn't infer protected attributes from other columns. You supply group data you're entitled to use when you run the audit:

audit = client.fairness.audit(
    model_id,
    sensitive_attributes=["gender"],
    y_true=y_true,
    y_pred=y_pred,
    group_data={"gender": genders},
)

If you don't have that data, the audit can't be run, and the tab says so. A failed audit with high severity can draft an entry in the AI Risk Register automatically.

For the metric definitions and thresholds, see Bias Detection.

Drift¶

Choose a drift type and a time range, then read the sub-tabs:

Sub-tab	Shows
Overview	The verdict and overall drift score for the period
Feature Analysis	Per-feature test results and which features drifted
Timeline	How drift has changed over time
Details	The test statistics behind each result

Refresh re-reads the latest results. Export Report downloads the drift results as CSV, and every chart can also be exported as PNG or CSV.

How drift is measured¶

Each feature in a recent window of predictions is compared with the model's baseline:

Continuous features: Kolmogorov–Smirnov test and Population Stability Index (PSI)
Categorical features: chi-squared test and Jensen–Shannon divergence
Concept drift: performance degradation against the baseline, and distribution-based methods including ADWIN

Sensitivity presets (low, medium, high) set all the thresholds together. See Drift Detection for configuration.

Keep a record¶

A quick check (client.drift.detect) answers has it drifted? without storing anything. A drift report (client.drift.create_report) is stored with an id you can cite, appears in the Model Register dossier, and can be turned into a Drift Report document in Reports. Use reports for anything you'll show someone else.

High-severity drift can draft an entry in the AI Risk Register automatically.

Who can do what¶

Action	Who
View results and verdicts	Everyone in the organization
Mark reviewed on a verdict	Analysts and administrators

Requesting an explanation, running a fairness audit or creating a drift report from the SDK, REST API or MCP server records one Governance Event. Doing it from the dashboard is free.