Model examples with Early Stopping breaking one-class-classification assumption?

[DeclanMcIntosh]

The YAML default configs for the models often use early stopping as below, based on image_AUROC or pixel_AUROC. Is this not using the labelled data and breaking the OCC assumption of the original method? Would this not make these methods rely on labelled test or validation data of both nominal and anomalous structures to replicate the results?

Example from examples/configs/model/reverse_distillation.yaml

trainer:
  callbacks:
    - class_path: lightning.pytorch.callbacks.EarlyStopping
      init_args:
        patience: 3
        monitor: pixel_AUROC
        mode: max
  check_val_every_n_epoch: 200
  max_epochs: 200