ANT: Angular-Norm Transport and Agentized Neural Computation [CPU-First]

ANT is not a new Taylor expansion. It is a mobile, route-aware and reliability-gated architecture for transporting and materializing local Taylor approximants. ANT является единым агентным протоколом для измерения, разделения, маршрутизации, проверки, реконструкции и сохранения информации в нейронных системах. ANT (Angular-Norm Transport) v1.4 presents a unified framework for representing neural credit assignment, routing, sparse memory, attention, model compression, and knowledge transfer as the movement of finite-lifetime, provenance-carrying information agents. The mathematical core separates every nonzero carrier into direction and log-magnitude, g=exp⁡(λ)ug=\exp(\lambda)ug=exp(λ)u, and derives exact polar reverse-mode identities. The framework then introduces a “One Reality, Two Time Scales” principle: one forward evaluation creates a shared residual field, scouts measure it, and many harvesters partition and transport the already observed information without an objective evaluation per agent. Complete residual partition is shown to reproduce a single vector–Jacobian product exactly, while selective routing changes the inductive bias and must be validated independently. Version 1.4 integrates eleven retained experiment families. These cover deceptive or dead gradients, partial observability, sphere and Stiefel transport, shared-reality reconstruction, program-carrying shortcut routes, first- and second-order Taylor packets, sparse spherical trail models, continual birth and evaporation, multi-metric attention, black-box oracle digestion, white-box structural digestion, and spectral/phasor trail compression. A major addition is the distinction between functional, structural, and hybrid digestion. In white-box digestion, every dense or convolutional teacher matrix is opened and rewritten as a sum of spherical rank-one trails. Full-rank trail networks reproduce trained MLP and CNN teachers to floating-point tolerance. At reduced rank, exact SVD remains the optimal Frobenius approximation, while sentinel-guided evaporation can deliberately remove high-energy but harmful watermark routes. On the Digits benchmark, a 75% rank trail model changes MLP clean/flipped accuracy from 0.903/0.217 to 0.881/0.917 and CNN accuracy from 0.578/0.450 to 0.867/0.867, using labelled counterfactual sentinel data. The spectral addon tests whether trails can be stored as complex phasors or sparse Fourier coefficients. Exact phase composition, Fourier phase-ramp shifts, and Parseval equivalence are numerically verified. The experiments also establish an important negative result: an FFT is a unitary change of coordinates, not automatic compression. Harmonic routes compress strongly, but the real singular trails tested here are not sufficiently sparse in a fixed Fourier basis. Scalar int4 direction quantization provides a better rate–distortion trade-off than naive sparse FFT/DCT or phase-only coding on these models. ANT is not claimed as a universal replacement for backpropagation, dense attention, SVD, low-rank models, or conventional distillation. The paper’s defensible contribution is a common, falsifiable protocol for measuring one reality, transporting typed information by alternative routes, validating what becomes persistent memory, reconstructing predictions from sparse trails, and selectively ingesting trained models. The archive contains the 53-page preprint, complete LaTeX source, a 12-question PROMPTS.md reading guide, claims and experiment manifests, eleven companion experiment packages, raw results, plots, trained toy checkpoints, and integrity hashes. S.V.E. Meta-License v4.0 [CC BY-NC-SA 4.0 + Symmetric Addendum]