aethelgard-ai 1.0.0

Python SDK for the Aethelgard AI Compliance & Fairness Verification Engine — formal proofs, not statistical guesses.

Aethelgard Python SDK

Enterprise AI Compliance & Fairness Verification — Formal Proofs, Not Statistical Guesses.

Aethelgard is the only AI governance engine that uses Z3 SMT formal verification to mathematically prove fairness compliance — not merely estimate it statistically. This SDK provides a Python client for the Aethelgard Engine API.

Why Aethelgard?

Capability	Fairlearn / AIF360	Aethelgard
Evidence Type	Statistical p-values	Z3 formal proofs (deterministic)
Intersectional Analysis	Manual specification	Exhaustive automatic discovery
Regulations	None built-in	6 frameworks (LL144, EU AI Act, SR 11-7, Colorado, FDA, Texas TRAIGA)
Confidence Intervals	Wilson CI only	Wilson CI + Z3 uncertainty-aware verification
Remediation	None	Automated minimum-cost remediation plans
Monitoring	None	Real-time drift detection with breach forecasting
Legal Protection	None	Attorney-client privilege safe harbor
Throughput	~100 rec/s (Python)	1.23 billion rec/s (Rust engine)

Installation

pip install aethelgard

For benchmarking utilities:

pip install aethelgard[benchmarks]

Quick Start

from aethelgard import AethelgardClient

client = AethelgardClient(
    base_url="http://localhost:8080",
    api_key="your-api-key",
)

# Run a bias audit (NYC LL144 / EEOC Four-Fifths Rule)
result = client.audit_ll144(
    model_name="hiring-model-v2",
    groups=[
        {"category": {"category_type": "Race", "value": "White"}, "total_candidates": 1000, "selected_candidates": 600},
        {"category": {"category_type": "Race", "value": "Black"}, "total_candidates": 800, "selected_candidates": 440},
        {"category": {"category_type": "Race", "value": "Hispanic"}, "total_candidates": 600, "selected_candidates": 300},
    ],
)

for group in result["group_results"]:
    print(f"{group['group']:20s}  IR={group['impact_ratio']:.4f}  Z3={group['z3_verified']}")

Supported Regulations

Regulation	Method	Jurisdiction
NYC Local Law 144	client.audit_ll144()	New York City
EU AI Act (Art 9-15)	client.audit_eu_ai_act()	European Union
Federal Reserve SR 11-7	client.audit_sr117()	US Banking
Colorado AI Act (SB 205)	client.audit_colorado()	Colorado
FDA 21 CFR 820	client.audit_fda()	US Medical Devices
Texas TRAIGA (HB 149)	client.audit_texas_traiga()	Texas

Key Features

Formal Verification (Z3 SMT Proofs)

Every audit produces machine-checkable mathematical proofs — not statistical estimates.

Intersectional Analysis

Automatically discovers and tests all combinatorial demographic intersections.

Multi-Regulation Crosswalk

Map compliance findings across frameworks (e.g., LL144 → EU AI Act → NIST AI RMF).

Zero-Knowledge Proofs

Prove fairness compliance without revealing model weights or training data.

Continuous Monitoring

Real-time drift detection with confidence-weighted alerts and breach forecasting.

API Reference

See the full documentation.

Benchmarks

Reproducible benchmarks comparing Aethelgard to Fairlearn, AIF360, and manual auditing across 10 real-world datasets (40K+ to 2.5M+ records):

pip install aethelgard[benchmarks]
python -m aethelgard.benchmarks

Citation

@software{aethelgard2026,
  author = {Ashuraliyev, Abduxoliq},
  title = {Aethelgard: Formal Verification for AI Fairness Compliance},
  year = {2026},
  url = {https://github.com/ProgrmerJack/Aethelgard},
}

License

Apache License 2.0. See LICENSE.