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 2.0 (Angular-Norm Transport and Agentized Neural Computation) presents a unified framework for treating neural credit assignment, routing, sparse memory, model compression, and persistent knowledge as the movement of finite-lifetime, provenance-carrying information agents. A nonzero carrier is separated into spherical direction and logarithmic magnitude, but the framework extends beyond polar gradients: agents share one measured residual field, construct typed shortcut routes, carry Taylor information, select metric-specific memories, digest frozen neural networks, and retain only structures that survive independent sentinel validation. The central extension in version 2.0 is a transition from storing visited states to storing validated generators of transport. Routes on a sphere are represented in the unit tangent bundle and may be encoded as rank-two Lie generators, block-phasor programs, piecewise geodesics, harmonic controllers, shared metric charts, or sparse ordered words over a colony-wide Queen codebook. The Queen stores reusable generators; each ant stores a compact genome containing generator identifiers, amplitudes, chronology or timing, reliability, provenance, and a minimal residual correction. Simultaneous Lie-algebra flow is explicitly separated from chronological Lie-group composition, and short codecs are accepted only when they preserve route geometry, endpoint, length, task output, and provenance. The 15 retained experiment families include shared-field harvesting, partial-observability search, true manifold transport, one-reality Fourier partition, dynamic shortcut mosaics, Taylor packets, sparse trail models, integrated continual colonies, multi-metric attention, black-box oracle digestion, white-box structural digestion, spectral and phasor codecs, native-geometry and shared-chart studies, tangent-bundle route compression, and Queen-codebook sparse genome experiments. New results show that independent planar rotations admit 512x phase compression while generic orthogonal operators do not; metric-native hyperbolic coding preserves Poincare geometry better than Cartesian quantization; shared charts can amortize over a colony; and activation-weighted full-SPD digestion can preserve neural function better than Frobenius-optimal SVD at the same rank. On three Digits MLPs, full-SPD rank-75% int8 digestion reaches median accuracy 0.9667, teacher agreement 0.9911, and 3.45x compression. Structured spherical routes are compressed by 107x-765x at near-machine precision, whereas a random route permits only about a 4x faithful fallback. A four-generator Queen recipe occupies 335 bytes instead of a 16,512-byte raw path, and recipe cost approaches about 209x compression after codebook amortization. Ordered generator words can be recovered exactly in a controlled test, while simultaneous mixture discovery remains a documented bottleneck. ANT 2.0 is not claimed as a universal replacement for backpropagation, dense attention, SVD, conventional distillation, or GPU execution. Exact gradients remain preferable when trustworthy; FFT and complex notation do not automatically compress arbitrary directions; dense charts and generator dictionaries have real storage and discovery costs; and public CIFAR/DeepSeek model-zoo evaluation remains an external reproducibility target where completed score files are not yet part of this record. The archive contains the complete preprint and LaTeX source, figures, a 12-question PROMPTS.md guide, claims and experiment manifests, publication metadata, integrity hashes, and 16 companion experiment or runner packages with code, raw results, tables, plots, and documented negative controls. S.V.E. Meta-License v4.0 [CC BY-NC-SA 4.0 + Symmetric Addendum] ANT_Preprint_v2_0_Final_Candidate.zip; Experiments:ANT_autonomous_native_metric_experiments_v0_1.zipANT_final_integrated_experiments_v1_0.zipANT_manifold_experiment_v0_1.zipANT_multimetric_attention_experiment_v0_1.zipANT_native_geometry_model_zoo_addon_v0_2.zipANT_oracle_digestion_digits_v1_0.zipANT_public_cifar_zoo_runner_v4_1.zipANT_queen_codebook_experiment_v0_1.zipANT_shared_reality_experiment_v0_3.zipANT_shortcut_mosaic_experiment_v0_2.zipANT_sparse_trail_model_v0_1.zipANT_spectral_phasor_addon_v0_1.zipANT_tangent_path_experiment_v0_1.zipANT_taylor_shortcut_experiment_v0_1.zipANT_v1_2_foundation_package.zipANT_whitebox_digestion_digits_v1_0.zip