Test AI system robustness and produce audit-ready documentation.

Robustness Testing

Boundary Value Testing

def test_edge_cases(model, feature_ranges):
    test_cases = []
    
    for feature, (min_val, max_val) in feature_ranges.items():
        # Minimum values
        test_cases.append({"description": f"Min {feature}", 
                           feature: min_val})
        # Maximum values
        test_cases.append({"description": f"Max {feature}",
                           feature: max_val})
        # Boundary values
        test_cases.append({"description": f"Near-min {feature}",
                           feature: min_val + 0.001})
        # Extreme outliers
        test_cases.append({"description": f"Extreme {feature}",
                           feature: max_val * 10})
    
    results = []
    for case in test_cases:
        try:
            prediction = model.predict([case])
            results.append({
                "case": case["description"],
                "prediction": prediction[0],
                "error": None
            })
        except Exception as e:
            results.append({
                "case": case["description"],
                "prediction": None,
                "error": str(e)
            })
    
    return results

Missing Data Robustness

import numpy as np

def test_missing_data_handling(model, X_test):
    results = {}
    
    for feature in X_test.columns:
        X_missing = X_test.copy()
        X_missing[feature] = np.nan  # Inject missing values
        
        try:
            predictions = model.predict(X_missing)
            results[feature] = {"handled": True, "predictions": predictions}
        except Exception as e:
            results[feature] = {"handled": False, "error": str(e)}
    
    missing_handled = sum(1 for r in results.values() if r["handled"])
    print(f"Missing data handled for {missing_handled}/{len(results)} features")
    return results

Audit Documentation

Model Card Generation

def generate_model_card(model, eval_data, audit_results):
    return {
        "model_overview": {
            "name": model.name,
            "version": model.version,
            "type": model.__class__.__name__,
            "trained_on": model.training_date,
            "intended_use": model.description
        },
        "performance_metrics": {
            "overall_accuracy": audit_results["accuracy"],
            "precision": audit_results["precision"],
            "recall": audit_results["recall"],
            "f1_score": audit_results["f1"]
        },
        "fairness_evaluation": {
            "groups_tested": audit_results["groups"],
            "disparate_impact_ratios": audit_results["fairness_ratios"],
            "passes_80_pct_rule": all(
                v >= 0.8 for v in audit_results["fairness_ratios"].values()
            )
        },
        "limitations": model.known_limitations,
        "ethical_considerations": model.ethical_notes,
        "audit_date": datetime.utcnow().isoformat(),
        "auditor": "AI Governance Team"
    }