Swiss Federal Supreme Court rulings already carry an official label. Apertus 1.5 8B reads the facts in German, French, or Italian. A small sheaf head on those hidden states returns a closed label. It does not write a judgment.
The facts of a case go in. The head returns one of two decisions. The first is the legal sub-area, one of thirteen official names. The second is the outcome, dismissal or approval. The label is the court's own. The reasoning section stays out of the text. Training uses the public sets rcds/swiss_judgment_prediction and rcds/swiss_law_area_prediction, both under CC BY-SA 4.0. Apertus stays frozen. Only the head is trained, for one pass over 63,328 original-language rulings.
On the official test file the legal-area head is right on 88.3 percent of 7,577 cases. Always choosing the most common sub-area is right on 26.0 percent. When the head is at least 90 percent confident it answers 72.5 percent of cases and is right on 98.0 percent of that slice. Rental and lease, employment, substantive criminal law, and urban planning are the strong classes. Public administration is weak, and some tax cases are read as criminal procedure. The validation file shows the same pattern: 93.4 percent accuracy against a 30.6 percent majority baseline.
The outcome task does not beat the obvious baseline. Test accuracy is 80.9 percent, the same as always saying dismissal. The head recalls 1.5 percent of approvals. A confidence of 90 percent or higher is always the label dismissal, including cases that were approved. That score is not a decision.
The useful piece is a triage aid for the legal area, shown only when the head is confident, with a person on everything else. It is not a system that decides a case.
