All documentation
Documentation
How Clause50 works, what it records, and what it produces.
Clause50 is an EU AI Act evidence platform. It classifies an AI system against the whole Act from one intake — prohibited practices under Article 5 are refused, high-risk systems are screened against all eight Annex III areas and documented across the nine sections of the Annex IV technical file, and the four Article 50 transparency duties are evidenced — and it answers every obligation that applies with append-only, Ed25519-signed evidence that anyone can verify without an account. It records evidence against the 2027 high-risk obligations today, so the file exists on the day they bind — or on the day a significant change brings a legacy system into scope. It pre-fills the EU declaration of conformity and refuses to sign it. The same intake also screens the system against the GDPR duties an AI system that processes personal data owes — the data protection impact assessment, Article 22 automated decisions, notices to the people whose data it processes, its Article 30 record entry and transfers — scored separately from the AI Act. It is not the quality management system, conformity assessment, EU database registration or the fundamental rights impact assessment — those stay yours, whichever tool you use.
Clause50 is not legal advice. For a plain-English guide to the regulation itself, rather than to this product, read the EU AI Act guide.
More detail
Start here
What Clause50 is — an EU AI Act evidence platform — and how to get your first AI system classified and evidenced.
Getting started
Sign in, name your organisation, and create your first AI system — the five minutes of setup before the real work begins.
Who it is for: Anyone who has just signed in and has nothing set up yet.
1. Define
Describe the AI system and see which transparency duties apply to it — the same first step as Define in the app.
Systems and intake
What an AI system record holds, where the intake questions come from, and what versioning means for evidence you have already recorded.
Who it is for: Whoever owns the description of the AI system — usually a product or engineering lead.
Coverage
What the score means, how gaps are counted, why upcoming obligations are excluded, and how it keeps itself up to date without you.
Who it is for: Whoever is asked how far along you are.
2. Prove
Record proof that you meet those duties, supplied by a person or fed in automatically by a machine.
Evidence
The four ways evidence arrives — a person states or uploads it, a connector or your own API pushes it — when to use each, why the record can never be edited or deleted, and how retraction lets you void a mistaken entry without erasing it.
Who it is for: Whoever will actually record proof: compliance, engineering, or both.
Connectors
The five connector types and what each one needs, getting an OpenAI or Anthropic Admin API key and a GitHub token, what each connector collects, the manual import format, and how hourly syncing works.
Who it is for: Whoever holds the OpenAI, Anthropic or GitHub credentials.
API
Push evidence in from your own CI, logs, or scripts with a per-system key: what it can and cannot do to your score, idempotency, and copy-paste examples.
Who it is for: Whoever owns the CI pipeline or the service that already produces the logs.
3. Generate
Produce the signed evidence pack and public disclosure text, and check that a pack is authentic.
Artifacts
Article 50 evidence packs and Annex IV technical files: draft to final, choosing the activity window, what ships inside the download, what the signature does and does not prove, and free preview versus paid download.
Who it is for: Whoever has to hand something to an auditor, a customer, or an authority.
Disclosure text
Generate the exact sentence your product should show for an Article 50 duty, in any of the 24 official EU languages, and why generating it does not by itself change your coverage score.
Who it is for: Whoever writes the words your users actually see in the product.
Verification
How to verify that an evidence pack is authentic and unchanged, what the signature proves and does not prove, and how to report a verification failure.
Who it is for: Anyone who has received a pack and needs to check it is genuine.
Reference
Everything that is not part of the three steps: notifications and short answers.
Alerts
The four kinds of alert in plain words, what the hourly pass actually does, and why an alert is a notification rather than a task.
Who it is for: Whoever needs to notice when something drifts.
FAQ
Short answers to the questions people hit in their first ten minutes.
Who it is for: Anyone who is stuck right now.
If you are stuck right now
Three questions account for most of the trouble people hit in their first ten minutes. Each has a direct answer:
- Why is my normal OpenAI or Anthropic key rejected?
- What is a GitHub access token, and how do I create one?
- What is the difference between recording a statement and uploading a document?
Everything else is in the FAQ.