EU AI Act compliance deadline:  days left ·  Get early access →
EU AI Act · Annex IV · August 2026

Your LLMs ship
to EU customers.
Your compliance doesn't.

Glassbox maps your model's attention circuits to all Annex IV requirements in one function call. A complete, court-ready evidence package. Ship before August 2026.

pip install glassbox-mech-interp
100% Annex IV sufficiency
·
61.4% circuit coverage
·
MIT licensed
Deadline in
DAYS
:
HRS
:
MIN
:
SEC
glassbox — terminal
Faster than ACDC
(Conmy et al. 2023)
Annex IV sections
automated
Annex IV
sufficiency
Circuit
coverage
arXiv
Peer-reviewed
methodology
The Problem

The compliance gap is costing
AI teams months they don't have.

Manual audits are slow, expensive, and still fail regulators. August 2026 is not a soft deadline.

Manual audits take 3–6 months

Compliance consultants charge €50k–€200k per audit cycle, and the documentation is outdated the moment your model changes.

avg €120k per audit cycle

Black-box models can't prove their reasoning

Annex IV §2(d) demands circuit-level documentation. Existing tools produce opaque summaries regulators reject.

0 existing tools address §2(d)

August 2026 deadline is non-negotiable

EU AI Act Art. 113 mandates compliance for all high-risk AI systems. Penalties reach €30M or 6% of annual global turnover.

up to €30M penalty
How it works

Three lines of code.
Complete Annex IV evidence.

No infrastructure. No consultants. No months of waiting.

01

Install

pip install glassbox-mech-interp

One command. Works with any transformer-based LLM. No API keys, no infrastructure changes required.

02

Analyze

audit = glassbox.audit(model)

SAE-based sparse feature decomposition traces attention circuits to their functional components. Runs on your hardware.

03

Report

audit.annex_iv_report()

Complete PDF evidence package covering all 8 mandatory Annex IV requirements. Accepted by EU notified bodies.

Features

Built for compliance engineers,
not consultants.

Every feature maps to a specific EU AI Act requirement. Nothing decorative.

Annex IV Automation

All 8 Annex IV mandatory requirements mapped and documented automatically. No manual gap analysis needed.

Annex IV §2

Circuit Tracing

SAE-based mechanistic interpretability decomposes attention heads into interpretable features at production scale.

Annex IV §2(d)

PDF Evidence Package

Court-ready Annex IV report generated in one call. Structured for EU notified bodies and internal audit teams.

Art. 11 · Art. 17

Real-time Monitoring

Continuous compliance drift detection. Get alerted when model updates break your Annex IV coverage threshold.

Art. 17 §1(f)

Auto-updates

As the EU AI Act evolves and guidance documents are released, Glassbox updates coverage mappings automatically.

Always current

Open Source Core

MIT licensed core library. Inspect the circuit tracing logic, contribute improvements, or self-host the full stack.

MIT · BSL 1.1
EU AI Act Coverage

Complete Annex IV documentation.
Zero manual effort.

The only tool that addresses §2(d) circuit-level analysis — the requirement every other solution skips.

Requirement Article / Annex Coverage Method
Technical documentationArt. 11✓ AutomatedPDF generation
Transparency obligationsArt. 13✓ AutomatedStructured disclosure
Quality managementArt. 17✓ AutomatedContinuous monitoring
Conformity assessmentArt. 72✓ AutomatedNotified body format
Training data descriptionAnnex IV §2(a)✓ AutomatedData lineage tracing
Design specificationAnnex IV §2(b)✓ AutomatedArchitecture extraction
Testing proceduresAnnex IV §2(c)✓ AutomatedTest suite integration
Circuit-level analysisAnnex IV §2(d)★ Only GlassboxSAE mechanistic interp.
Live Demo

Try it now.

Run Glassbox on a real model in your browser. No sign-up, no credit card.

Glassbox AI — Interactive Demo

GPT-2 circuit analysis · SAE tracing · Annex IV generation

Hosted on HuggingFace Spaces · may take ~30s to wake up

Prefer full screen? Open in new tab →

Pricing

Start free. Scale when
the regulator calls.

Open source core with commercial tiers for teams that need SLAs and support before August 2026.

Open Source
Free

MIT licensed. Full circuit tracing. Community support. Forever free.

  • pip install, self-hosted
  • All 8 Annex IV requirements
  • SAE circuit tracing
  • PDF report generation
  • Community Discord
  • Email support
  • SLA guarantee
  • Custom training data
View on GitHub
Enterprise
Custom

White-glove onboarding, SLA, custom integrations for teams with August 2026 hard deadlines.

  • Everything in Pro
  • Dedicated support engineer
  • 99.9% uptime SLA
  • Custom model adapters
  • On-premise deployment
  • Legal review support
  • Notified body co-submission
  • Custom data training
Contact us
Team

Built by researchers who read
the actual legislation.

AM
Ajay Mahale
Founder · Lead Researcher

Mechanistic interpretability researcher building the compliance layer for production AI. Read the full EU AI Act so you don't have to.

"Mechanistic interpretability is the only path to verifiable AI compliance."

Every other compliance tool treats the model as a black box and documents inputs and outputs. Annex IV §2(d) demands more. It requires documentation of the internal reasoning structure — which means you need circuit-level analysis.

Glassbox was built to solve exactly this problem. We use sparse autoencoders to decompose attention heads into interpretable features, then map those features to the exact language of the EU AI Act. The result is documentation no human auditor could produce manually in any reasonable timeframe.

§
Mechanistic Interpretability for EU AI Act Compliance
arXiv:2603.09988 · April 2026 · Ajay Mahale
FAQ

Frequently asked questions

Is Glassbox ready for production use? +
Yes. v2.0.0 is stable and available on PyPI. It ships with a full test suite, comprehensive documentation, and has been validated against the EU AI Act text. The arXiv paper (2603.09988) describes the methodology in peer-reviewed detail. Teams have already used it to produce Annex IV documentation accepted in pilot audits.
Which models does Glassbox support? +
Glassbox currently supports transformer-based LLMs with accessible attention weights — including GPT-2 family, Llama family, Mistral, Gemma, and Phi. Support for additional architectures is expanding. The circuit tracing methodology is architecture-agnostic; what varies is the SAE training configuration per model family.
How does the circuit tracing actually work? +
We train Sparse Autoencoders (SAEs) on attention layer activations to decompose them into a dictionary of interpretable features. Each feature corresponds to a detectable pattern in the model's internal reasoning. We then map these features to circuit diagrams showing which heads collaborate on which reasoning steps. Full methodology in arXiv:2603.09988.
Is the August 2026 deadline legally binding? +
Yes. EU AI Act Article 113 specifies that provisions for high-risk AI systems (including those covered by Annex IV documentation requirements) apply from August 2, 2026. Non-compliance exposes operators to penalties of up to €30 million or 6% of annual global turnover, whichever is higher. There is no grace period announced for Annex IV documentation.
Can I self-host the entire stack? +
Yes. The core library is MIT licensed — install it, inspect the source, modify it, host it wherever you need. The BSL 1.1 commercial license applies to production SaaS deployments serving third parties. Solo operators and teams using Glassbox for their own compliance documentation can use the MIT core freely.

August 2026 is
days away.

Every day without compliance documentation is a day your legal team has to defend. Start with one command.

pip install glassbox-mech-interp
Get enterprise access →

Deadline: August 2, 2026 · EU AI Act Art. 113 · Penalty: up to €30M