When a teacher asks for an advisory assessment of a finding, the extract goes to a language model we run ourselves on a GPU server in the EU. LUCA, the study model, is built on the same footing and is still in development. There is no API call to OpenAI, Google or any other service outside Europe.
- No third party
- Open weights on our own machine. Student data never reaches an external AI service, and there is no third-country transfer to assess.
- Advisory, never decisive
- The model’s answer is marked as advisory and attached to the finding as an aid for the teacher. No student is flagged or judged by a model.
- If it fails, it does not guess
- If the model does not answer, the question is noted as unanswered in the summary. The rules keep running without it, and the exam is unaffected.
Some school tools run on OpenAI's GPT-4o and GPT-5.1. Even when the archive itself is stored in the EU, the conversation with the model is sent to OpenAI's servers in the US. That does not apply to our own models.