Scope
The paper applies the relative mapping objects, transversality results, and derived intersection models of Differential Topology in Higher Categories to smooth parametrized neural networks. It also builds on Representation Theory of Parametrized Neural Models on Spaces.
Readouts and decision boundaries
An affine readout of an immersed smooth feature map is submersive on first jets. Free final biases give generic transverse classification boundaries even when the feature map is not immersive. A quadratic circle classifier gives an explicit fold discriminant. A different exact circle network has regular signed boundary points with cyclic monodromy in the parameters. Its persistence over the parameter circle retains transport data that its individual fiber barcodes forget.
Symmetry and reconstruction
At a fixed parameter, its stabilizer representation can rule out surjectivity of the normal derivative. A reflected torus with two sign outputs exhibits a derived obstruction line, even though a transforming bias makes the total zero equation regular. Smooth exact ambient reconstruction through a fixed Euclidean encoder is possible precisely when the encoder embeds the compact data manifold. For the unit sphere and continuous ambient decoders with latent width at most the sphere dimension, the optimal uniform error is one.
Evidence and authorship
The manuscript gives proofs and explicit worked models. It is submitted under an anonymous byline and was prepared with AI assistance in drafting, typesetting, and checking. The attached LaTeX file contains the entire manuscript and is designated for text indexing upon publication.