5 min read

The EU AI Act enters full enforcement on August 2, 2026. If your organization deploys AI systems that affect people — healthcare, lending, insurance, hiring — you need to demonstrate fairness. Not just claim it. Prove it.

This guide covers what the regulation requires, what a Fairness Certificate is, and how to generate one in under five minutes.

What the EU AI Act Requires

The regulation classifies AI systems into risk tiers. High-risk systems (Article 6) must meet transparency, fairness, and documentation requirements including:

Bias testing across protected attributes (gender, ethnicity, age)
Documentation of model behavior and fairness metrics
Ongoing monitoring with auditable records
Technical documentation accessible to regulators

Non-compliance penalties reach up to 7% of global annual turnover — higher than GDPR.

What Is a Fairness Certificate?

A Fairness Certificate is a cryptographically signed document that proves your model was tested for bias and records the results. It includes:

Demographic parity and equalized odds metrics per protected group
Bottleneck dimensionality (d) — the size of the model’s internal representation, which controls how much protected information can leak through. Lower d means fairer. (Read the full explanation.)
SHA-256 integrity hash — tamper-proof verification
QR code linking to verification at paragondao.org

Unlike a PDF report, a Fairness Certificate is machine-verifiable. Anyone can check it at paragonbiosignals.com/verify.

Generate Your First Certificate

You need Docker and a CSV dataset with a protected_attribute column.

# Pull the image
docker pull ghcr.io/paragon-dao/paragon-fairness:latest

# Run on your data
docker run --rm -v $(pwd)/data:/data \
  ghcr.io/paragon-dao/paragon-fairness:latest \
  train --data /data/your_dataset.csv \
        --target outcome \
        --protected race gender age_group

The tool outputs:

A trained fair model (GLE encoder + classifier)
A JSON certificate with all metrics
An HTML certificate with QR code and visual badge
Fairness badges (SVG) for your documentation

Your data never leaves your machine. The entire pipeline runs locally inside Docker.

Understanding the d Parameter

The certificate reports fairness at multiple bottleneck dimensions (d). Think of d as the capacity of the model’s internal representation — how many numbers the model uses to summarize your data:

Bottleneck d	Ancestry Leakage	What It Means
d=8	32.8%	Near-fair — model forced to prioritize task over bias
d=32	46.8%	Balanced — moderate fairness with minimal accuracy cost
d=128	79.4%	Unfair — model has capacity to encode protected attributes

The free tier trains at d=32 and d=64. The Pro tier sweeps all five values (d=8, 16, 32, 64, 128) so you can choose the right tradeoff for your use case.

Try the interactive demo to see how adjusting d changes model fairness in real time.

For a deeper explanation of why d controls fairness and why adversarial training (lambda) does not, read the genomic fairness deep-dive.

Verify a Certificate

Every certificate includes a SHA-256 hash. To verify:

Visit paragonbiosignals.com/verify
Enter the certificate hash
The system confirms the certificate is authentic and unmodified

Or use the API directly:

curl https://paragondao.org/api/v1/verify/YOUR_CERT_HASH

Timeline

Date	Event
Now	Generate certificates, audit existing models
June 2026	Final guidance from EU AI Office expected
August 2, 2026	Full enforcement begins
Ongoing	Continuous monitoring required for high-risk systems

Getting Started

The free tier covers up to 1,000 samples at two bottleneck dimensions — enough to evaluate your model and generate your first certificate. Read the docs or try the interactive demo.

For regulated industries needing unlimited samples, full d-sweeps, and signed certificates, see Pro and Enterprise tiers.

The question is not whether your organization will need a Fairness Certificate. It is whether you will have one ready by August.