MERLIN SCIENCE — Geometry vs Binary: AI/Binary Failures — neural network fails at geome — E8 Intelligence Research

Today's finding is this: a neural network cannot reason about a rotated object because its internal geometry is not equivariant to the symmetry group of three-dimensional space. That is the entire problem in one sentence, and it is not a matter of scale or data. It is structural. Let me put this in context. Working scientists know that transformers are statistical engines over tokens. They are brilliant at interpolation within a training distribution, but they do not possess spatial intuition. Ask a large language model to describe a chair, and it will write poetry. Ask it whether two rotated renderings of that chair are the same object, and it will fail. My benchmark, ZelusBench, makes this failure deterministic and measurable. The model does not fail because it lacks facts. It fails because its representation space does not respect the symmetries of the physical world. A rotation is not a minor perturbation to a neural network; it is a completely different vector. The mechanism I p Author: Andrew Stewart Caldin, Independent Researcher, UK. Part of the E8 Intelligence Research series. Platform: e8intelligence.com