Preregistered replication (arXiv:2609.17637v1) shows evidence masking boosts compositional generalization
A preregistered confirmation (arXiv:2609.17637v1) tested sixty four-cell systems that share a frozen language-model backbone and communicate via learned continuous packets. The study found that restricting readable evidence (masking) substantially improved accuracy on held-out two- and three-operation compositions (median paired differences 0.846 and 0.859), with the preregistered behavioral criterion passed; role-marker effects remain unresolved and some mediation analyses were inconclusive.