Feature overview¶
Which guide do I need?¶
graph TB
Start[What do you need?]
Start --> Dev{Implementing?}
Dev -->|Yes| DevType{Your focus?}
DevType -->|CI/CD| MLEng[ML Engineer Guide]
DevType -->|Notebooks| DataSci[Quickstart + Custom Data]
DevType -->|Validation| Validator[Validator Guide]
Dev -->|No| Role{Your role?}
Role -->|Compliance| CompOff[Compliance Officer Guide]
Role -->|Industry needs| Industry{Industry?}
Industry -->|Banking| Banking[Banking + SR 11-7]
Industry -->|Insurance| Insurance[Insurance Guide]
Industry -->|Healthcare| Healthcare[Healthcare Guide]
Industry -->|Fraud| Fraud[Fraud Detection]
Role -->|Exploring| Explore[Start with Quickstart]
style MLEng fill:#d4edda
style DataSci fill:#d4edda
style Validator fill:#d4edda
style CompOff fill:#d4edda
style Explore fill:#e1f5ff
Quick navigation:
- Implementing audits? → ML Engineer Workflow for CI/CD and debugging
- Exploring in notebooks? → Quick Start + Custom Data Guide
- Need evidence for regulators? → Compliance Officer Workflow
- Industry-specific guidance? → Banking | Insurance | Healthcare
- Independent verification? → Model Validator Workflow
Quick reference¶
GlassAlpha capabilities and where to learn more:
| Feature | Guide | Reference | Compliance |
|---|---|---|---|
| Group Fairness | Configuration | Fairness Metrics | SR 11-7 §V |
| Intersectional Fairness | Configuration | Fairness Metrics | SR 11-7 §V |
| Individual Fairness | Configuration | Fairness Metrics | SR 11-7 §V |
| Dataset Bias | Detection Guide | - | SR 11-7 §III.C.2 |
| Calibration | Configuration | Calibration Reference | SR 11-7 §III.B.2 |
| Robustness | Configuration | Robustness Reference | EU AI Act Art. 15 |
| Shift Testing | Shift Testing Guide | - | SR 11-7 §III.A.3 |
| Reason Codes | Reason Codes Guide | - | ECOA §701(d) |
| Preprocessing | Preprocessing Guide | - | SR 11-7 §III.C.1 |
Why choose GlassAlpha?¶
How GlassAlpha compares¶
| Feature | GlassAlpha | Fairlearn | AIF360 | Commercial Tools |
|---|---|---|---|---|
| Audit PDFs | ✅ Professional, deterministic | ❌ No reports | ❌ No reports | ✅ $$$ |
| Custom Data in 5 min | ✅ Yes | ⚠️ Complex setup | ⚠️ Complex setup | ⚠️ Support needed |
| Built-in Datasets | ✅ Multiple ready to use | ❌ None | ⚠️ Few | ✅ Limited |
| Model Support | ✅ XGBoost, LightGBM, sklearn | ⚠️ sklearn only | ⚠️ Limited | ✅ Varies |
| Deterministic Results | ✅ Byte-identical (same platform+Python) | ⚠️ Partial | ❌ No | ⚠️ Varies |
| Offline/Air-gapped | ✅ 100% offline | ✅ Yes | ✅ Yes | ❌ Requires internet |
| Cost | ✅ Free (Apache 2.0) | ✅ Free (MIT) | ✅ Free (Apache 2.0) | 💰 \(5K-\)50K+ |
| Regulatory Ready | ✅ Audit trails + manifests | ❌ No trails | ❌ No trails | ✅ $$$ |
| Learning Curve | ✅ Quick start available | ⚠️ Steep | ⚠️ Steep | ⚠️ Training needed |
Bottom line: GlassAlpha is the only OSS tool that combines professional audit PDFs, easy custom data support, and complete regulatory compliance.
Designed for regulatory compliance¶
- Deterministic outputs - Byte-identical HTML reports on same platform+Python+config
- Complete lineage - Git SHA, config hash, data hash, seeds recorded
- Professional formatting - Publication-quality reports with visualizations
- Audit trails - Immutable run manifests for regulatory submission
Note: HTML reports are byte-identical across platforms when using the same Python version and config. PDFs may have minor visual differences across platforms due to font rendering and layout engine variations, but are suitable for human review and regulatory submission. See determinism guide for details.
On-premise first design¶
- No external dependencies - Runs completely offline
- File-based approach - No databases or complex infrastructure needed
- Full reproducibility - Immutable run manifests for audit trails
- Air-gapped compatible - Works without internet access
See trust & deployment details →
Simplicity as a core principle¶
- Single command -
glassalpha audithandles everything - YAML configuration - Policy-as-code for compliance requirements
- Fast execution - 2-3 seconds for demo mode, 5-60 seconds for full audits
- Clear errors - Actionable messages with fix suggestions
Supported models¶
| Model Type | Status | Notes |
|---|---|---|
| XGBoost | Production | TreeSHAP integration optimized |
| LightGBM | Production | Native integration available |
| Logistic Regression | Production | Full scikit-learn compatibility |
Additional model types available through extension framework.
Example configuration¶
Working configuration structure:
# Direct configuration example
data:
path: data/german_credit_processed.csv
target_column: credit_risk
protected_attributes:
- gender
- age_group
- foreign_worker
model:
type: xgboost
params:
objective: binary:logistic
n_estimators: 100
max_depth: 5
explainers:
strategy: first_compatible
priority:
- treeshap
- kernelshap
metrics:
performance:
metrics:
- accuracy
- precision
- recall
- f1
- auc_roc
fairness:
metrics:
- demographic_parity
- equal_opportunity
reproducibility:
random_seed: 42
This configuration format supports deterministic, reproducible audits.
See full configuration reference →
Contributing¶
We welcome contributions to enhance GlassAlpha's capabilities:
Enhancement areas¶
- Additional models - Neural networks, time series, custom integrations
- Advanced explanations - Counterfactuals, gradient methods, interactive visuals
- Extended compliance - Additional frameworks, custom templates, industry metrics
- Performance - Large dataset optimization, parallel processing
- Documentation - Examples, tutorials, best practices