Question and framework
How do weights, feature representations, and observable laws vary when a neural model is defined locally over a space? A layer over an object $B$ of an infinity-topos is represented by its classifying map $P\to[X,Y]_B$. Reparametrizations are morphisms in the corresponding slice. Composition is induced by internal evaluation. The paper proves base change and descent along effective epimorphisms, then treats probability-valued outputs using strong monads in settings where the requisite structure is specified.
Statistical and geometric specializations
Smooth charts recover manifold-valued maps, latent manifold decoders, and conditional statistical models. The finite-output case derives the pullback Fisher tensor, its contraction under fixed stochastic postprocessing, and its degeneracy along unidentifiable directions. For compact group actions, the paper develops covariant kernels on homogeneous feature bundles and classifies bounded Hilbert-space layers linear in a finite-dimensional moving parameter. Tensor multiplicities determine which input and output types a layer can couple. A circle example realizes a learned Fourier shift.
Higher comparison
The quotient-stack model distinguishes equal observable laws from equivalent parameter presentations. A sign-symmetric decoder gives a smooth law on the circle with no global smooth scalar weight in the same architecture. The stable setting uses homotopy fixed points to compute equivariant mapping spectra and includes an explicit modular example of a higher intertwiner. The manuscript builds on Representation Theory in Higher Categories.
Evidence and authorship
The manuscript supports its claims through mathematical proofs and worked examples. It was prepared with AI assistance in drafting, typesetting, and checking the presentation, and is submitted under an anonymous byline. The separately attached LaTeX file is the complete source and is selected for indexing on publication.