A preprint reports higher VN-SST scores than a Transformer on three neural benchmarks, with the clearest gap under limited training data.