Apertus 8B is the reader. A sheaf head sits on its hidden states. A note and a closed question go in. The head returns a choice, a yes or no, or a score. It does not write an open answer.
The training notes are published datasets that already carry their own label: intents, topics, reviews, entailment, and the Federal Supreme Court tasks. A row stays only when that label still matches the stored text. Copies are removed, and one source cannot fill the file.
The court rulings are also a project of their own, Apertus Court Decisions, trained only on those rulings. This page is the general decision model, trained on the larger pack. The 8B stays frozen. The sheaf head is what gets trained. The 70B does not fit this machine. The head is not trained yet.
