Prior work
This survey follows Actions of Toposes on Toposes and distinguishes a geometric action of a monoidal Grothendieck topos from external group equivariance and internal group actions. It reviews neural-network semantics using topoi and stacks, relative classifying topoi for symbolic reasoning, active learning of nominal automata, cellular sheaf learning, groupoid-equivariant kernels, and categorical accounts of training and probabilistic inference. Direct learning applications of geometric actions by monoidal topoi were sparse in the sources reviewed.
Worked applications
A central action on a group classifying topos gives a criterion for when nuisance-invariant layers factor through input orbits. A finite example distinguishes orbit features from linear averaging, which can discard the label signal. A second example uses a C3 group sheaf with inversion monodromy around the circle. Its equivariant linear filter coefficients split into a trivial rank-two bundle and a real Möbius line. The paper computes the constraints on flat parameters and the effect of a hidden orientation in an exact regression model.
Evidence and authorship
The manuscript includes proofs, exact calculations, and a bibliography with links to prior work. It was prepared with AI assistance in drafting, typesetting, and checking, and is submitted anonymously. The attached standalone LaTeX contains the complete manuscript and is designated for indexing upon publication.