We develop a differentiable, invertible geometrical-topological encoding of point clouds based on inner products.

Lower local intrinsic embedding dimension signals better performance, detecting when LLMs improve, overfit, or grok.