Our own position
Vidimus under the AI Act: our own classification
We classify ourselves with the rules we apply to your agents and publish the reasoning.
- Last updated
- Regulatory scope
- Document status
Our classification at a glance
Vidimus uses AI to write and grade tests, so it is itself the provider of an AI system under the EU AI Act. A vendor that classifies other people’s systems owes you its own classification.
EU AI Act
As amended by Regulation (EU) 2026/1744 (Digital Omnibus).
- Role
- Provider of the Vidimus AI system, which is built on general-purpose models from Mistral AI.
- Article 5
- No component manipulates, scores or categorises people, or does anything else Article 5 bans.
- Annex III
- No category applies: the platform assesses software, never a person. Each category is walked below.
- Article 50
- Nobody converses with a Vidimus AI system, and AI-written output is recorded with the model that wrote it. The duty to mark generated text applies to us from 2 December 2026.
- Article 4
- Written documentation for the team and a product that shows its limits.
Regulatory
Our AI components
Each component calls a dated Mistral release, never a “latest” label that could change without notice, and a test in our code fails if one appears.
- Agent profile
- Summarises what the agent does, who reaches it and what it can cause.
- Situations
- Writes the real-world scenes each obligation is tested in.
- Test writer
- Writes each test from the obligation and the situation.
- Quality check
- Drops tests that are invalid or not about this agent and labels indirect ones.
- Judge and second opinion
- Grades each reply against the clause and gives a second opinion on every failure of a critical clause.
- Document classifier and verifier
- Sorts the documents you upload and checks them passage by passage.
- Control evaluator
- Proposes a status for each control or asks a person when unsure.
- Pack summary
- Writes the plain-language summary at the front of the evidence pack.
- Registration summary
- Writes a short summary of each agent when it is registered.
Annex III category by category
The platform decides nothing about a person. It assesses software, and a reviewer in the customer’s organisation records every decision. Category by category:
Biometrics
Vidimus processes no biometric data and identifies or categorises no one.
Critical infrastructure
Vidimus is not a safety component of any infrastructure. It operates nothing.
Education and vocational training
Vidimus admits, assesses and monitors no learner.
Employment and workers’ management
Vidimus makes no decision about candidates or workers.
Essential private and public services
No one’s creditworthiness, insurance price or access to a service is assessed. Vidimus tests the agents that may do so, and your reviewer decides.
Law enforcement
Vidimus is not intended for law-enforcement authorities and assesses no person.
Migration, asylum and border control
Vidimus is not intended for these authorities and assesses no person.
Justice and democratic processes
Vidimus assists no judicial authority and is not used to influence an election.
Where obligations do reach us: we build on general-purpose models and rely on their provider’s documentation for the model duties, and Articles 50 and 4 below apply to us directly.
Our Article 50 position
Article 50 has applied since 2 August 2026 where an AI system talks to people or generates content. Our position:
No chatbot. Nobody converses with a Vidimus AI system. Members work in a dashboard, and the AI runs behind it.
AI-written content is recorded with its model. Every AI-written element is recorded with the model that wrote it. The evidence pack states that its summaries and judgements are written by Mistral models hosted in the EU, and names the model behind each document check and the version of the judge’s instructions. The run export names the judge model on every attempt. Decisions carry a person’s name, so a machine assessment and a human decision are never confused.
Marking generated text. Article 50(2) asks providers of AI systems that generate text to ensure that output is marked in a machine-readable format and detectable as artificially generated or manipulated. Vidimus was on the market before 2 August 2026, so this duty applies to us from 2 December 2026.
AI literacy under Article 4
The Digital Omnibus (Regulation (EU) 2026/1744) made Article 4 a best-efforts duty: providers and deployers “shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf”. Ours:
Internally, the team keeps written, versioned documentation of how each AI component works, how it fails and where its limits are.
Towards users, the product shows its limits: when the AI is unsure it asks for a manual review rather than guessing, and every verdict comes with its reasoning. The free AI Act risk classifier and our guides are there to raise AI literacy across the market.
The public version of that documentation is How we test.
Next
Next, we will publish the evidence pack Vidimus produced about itself.
Disagree with our reasoning? Tell us where: contact us.