ANT 3.0: Angular-Norm Transport, Verified Whole-Object Digestion, and Living Colony 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 3.0 is an integrated research release on representing learning, routing, compression, persistent memory, and verified model transformation as the movement of finite-lifetime, provenance-carrying information agents. The framework begins with an exact angular/log-norm reparameterization of nonzero vectors and develops shared residual fields, scouts and harvesters, manifold transport, shortcut mosaics, sparse trail memories, spectral/native codecs, tangent-bundle route programs, and Queen codebooks of reusable generators. The v3.0 extension treats neural networks and graphs as typed causal multigraphs whose digestible units are whole functional organs rather than isolated tensors. Candidate detectors based on attention, activations, graph bottlenecks, curvature, or gradients are separated by a detection/selection firewall from causal auditors, atomic multi-site bite planners, rare/OOD Sentinels, named-shock fortifiers, and deterministic `.antcolony` serialization. A candidate is accepted only when it satisfies a declared quality contract together with real byte and latency accounting, checksum validation, and teacher-unloaded replay. Measured additions include: (i) a five-seed whole-organism Digits experiment with median 7.83× storage reduction and 1.86× measured speedup while preserving a rare subgroup that attention-only selection destroys; (ii) a planted shared rank-32 Queen runtime with 13.62–13.68× measured speedup when the native law exists; (iii) graph experiments separating bottleneck detection from functional selection, reconstructing repeated modules exactly at 6.04× transparent description reduction, and showing that maximizing algebraic connectivity need not repair a named rare function; and (iv) a 50-seed living-colony search experiment with an 88/10/2 risk prior, four-ant squads, decaying Historian memory, creator-built multi-site worlds, a budgeted private Queen gate, and a sealed final audit. Squads plus Historian recover the planted atomic synergy in all seeds, while more adaptive exploration is marginally worse than the fixed squad policy on this family. The release preserves negative results. The earlier Qwen2.5-3B pilot shows that independent aggressive low-rank replacement can produce apparent speedups while destroying perplexity and top-1 fidelity. The included Stage-III Qwen runner is therefore infrastructure and a self-tested protocol, not evidence of a successful large-language-model compression result. Likewise, the living-colony experiment is synthetic and does not establish a universal 88/10/2 optimum or mathematical antifragility. The package contains the 78-page PDF, full LaTeX source, all figures, reproducible code and raw results for the new experiment, a Stage-III Qwen runner, claims and limits, public-run status, study prompts, checksums, and prior companion experiment archives. S.V.E. Meta-License v4.0 [CC BY-NC-SA 4.0 + Symmetric Addendum] ANT_Preprint_v3_0_Final_Candidate.zip; Experiments:ANT_ariadne_labyrinth_addon_v0_1.zipANT_autonomous_native_metric_experiments_v0_1.zipANT_colony_format_addon_v0_1.zipANT_final_integrated_experiments_v1_0.zipANT_manifold_experiment_v0_1.zipANT_model_graph_digestion_v0_1.zipANT_multimetric_attention_experiment_v0_1.zipANT_oracle_digestion_digits_v1_0.zipANT_Preprint_v2_0_Final_Candidate.zipANT_queen_codebook_experiment_v0_1.zipANT_sparse_trail_model_v0_1.zipANT_spectral_phasor_addon_v0_1.zipANT_stage3_verified_qwen_digestion_v0_1.zipANT_tangent_path_experiment_v0_1.zipANT_taylor_shortcut_experiment_v0_1.zipANT_verified_smart_digestion_v0_1.zipANT_whitebox_digestion_digits_v1_0.zip