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) proposes an agent-based framework for neural optimization, credit assignment, sparse computation, and selective knowledge transfer. Instead of treating gradients as passive tensors confined to a fixed computational graph, ANT represents optimization information as finite-lifetime, provenance-carrying agents that can be measured, divided, routed, validated, and stored. The foundational representation separates a nonzero signal into direction and log-magnitude, [g=e^\lambda u,\qquad u\in S^{d-1},\qquad \lambda=\log|g|.] This decomposition does not eliminate analytic vanishing or exploding gradients; rather, it isolates dynamic range in the radial channel while preserving an explicit directional channel. The paper derives exact polar reverse mode and introduces reliability-gated combinations of exact, surrogate, structural, forward-mode, and zeroth-order information. A central principle is “One Reality, Two Time Scales.” At each macro-step, the model measures a shared residual or food field once. A small scout population evaluates informative directions, while a larger harvester population transports standardized portions of the already measured field without independently re-evaluating the objective. For the shared-field estimator, the aggregate is unbiased for the measured projected field, and its carrier mean-square error decreases as (O(M^{-1})) with the number of harvesters (M). The framework is extended beyond optimization to dynamic shortcut routing, residual mosaic reconstruction, Taylor packets carrying slope, curvature, and reliability, spherical sparse trail memories, metric-aware attention, continual route birth and evaporation, and selective oracle digestion. In ANT-Digestion, a frozen neural network is treated as a black-box environment: scouts sample its outputs and local sensitivities, while a sparse student retains only routes that survive sentinel validation and minimum-description-length penalties. The accompanying experiments include controlled optimization pathologies, partial-observability search, manifold transport, shared-reality reconstruction, shortcut routing, Taylor-aware transport, sparse trail compression, multi-metric attention, and black-box distillation of neural networks on handwritten-digit data. The results support the individual mechanisms at controlled scale while also identifying important negative results: curvature does not establish truth, pheromone can cause trail collapse without exploration, sparse routing does not automatically reduce score-computation cost, and ANT does not universally outperform reliable backpropagation, dense attention, or standard distillation. ANT is therefore presented not as a universal replacement for existing neural methods, but as a unified protocol for measuring, transporting, validating, reconstructing, and selectively preserving information in adaptive learning systems. This record contains the complete preprint, LaTeX source, experiment manifest, claims matrix, reproducibility code, raw results, aggregate tables, and companion technical notes. S.V.E. Meta-License v4.0 [CC BY-NC-SA 4.0 + Symmetric Addendum]