I'd like to see similar testing done comparing models where the "misaligned" data is present during training, as opposed to fine-tuning. That would be a much harder thing to pull off, though.