Diotima: Compliance-Grade AI Infrastructure for Formative Assessment in Education
A practical guide to designing EU AI Act Annex III compliance into educational AI from first principles, not retrofitting it after the fact.
What's inside
The EU AI Act classifies AI systems used to assess students in formal education as high-risk under Annex III. Most educational AI products treat compliance as a downstream documentation task. This paper argues that compliance-by-design is not a cost of participation in regulated education markets, it is the only credible path to adoption at scale.
Using the Diotima formative assessment platform as a case study, the paper presents a transferable framework for designing high-risk educational AI from regulatory first principles, with each governance dimension mapped to its specific EU AI Act obligation.
- How to scope the high-risk AI perimeter narrowly under Annex III so governance remains tractable
- How to ground generative outputs in approved curriculum materials to prevent hallucination and curriculum drift
- How to enforce Article 14 human oversight as a structural property of the workflow, not a policy aspiration
- How to differentiate explainability by stakeholder role (teachers see full reasoning, students see progressive rubric reveal)
- How to govern third-party foundation models with structured benchmark suites (MMLU, GPQA, LongFact, IFEval, BBQ, RealToxicityPrompts, SimpleSafetyTests)
- How to embed Article 72 post-market monitoring as a byproduct of normal operation rather than a separate workstream
Who it's for
Educational AI product teams, EdTech founders preparing for EU AI Act enforcement, compliance and risk officers at universities and schools, investors evaluating regulatory exposure in EdTech portfolios, and policy professionals working on educational AI governance.
Get the whitepaper
Fill in your details and we'll email you the PDF. We'll also let you know when the companion paper on operational metrics is published.