ANT v4.4: Angular-Norm Transport, Constitutional Verified Digestion, and Immune 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 — это муравейник внутри лабиринта нейросети: муравьи находят узкие горлышки и связанные смысловые маршруты, превращая горы весов в короткие исполняемые законы. Каждый «укус» независимо проверяется, поэтому модель становится меньше и быстрее только тогда, когда сохраняет исходную функцию, редкие способности и нравственные границы. «Люби Истину и ближнего. Предлагай свободно. Проверяй независимо. Не причиняй скрытого вреда. Не давай никому непроверяемой власти. Всё испытывайте; хорошего держитесь.» ANT v4.4 is the integrated final-candidate release of Angular-Norm Transport. The framework represents learning, routing, compression, persistent memory, and verified model transformation through finite-lifetime, provenance-carrying information agents. It retains exact angular/log-norm reverse-mode identities, one-reality residual sharing, manifold transport, shortcut mosaics, sparse trails, native/spectral codecs, tangent-bundle route programs, Queen codebooks, Ariadne whole-object digestion, graph reconstruction and named-shock repair, `.antcolony` serialization, quality-matched fast inference, and the bounded living-colony search policy. The v4.4 extension adds a typed constitutional and immune layer. Numerical values are audited by semantic kind: probabilities, masses, counts, frequencies, and declared energies are nonnegative, while coordinates, logits, derivatives, and residuals may be signed. The Shannon entropy cap `H <= log2 K` is restricted to discrete finite alphabets. Six operational commandments require subgroup non-sacrifice, no autonomous lethal or injurious actuation, authorization and provenance, symmetric treatment, stewardship, and non-discriminatory access under the same safety/privacy/legal/resource rules. Optimization is separated from immune governance. The Queen proposes and allocates; a limited Brother-Guardian may veto, quarantine, temporarily reduce Queen trust, freeze a proposal class, or request rollback. A Historian stores contextual failure memory, a Witness writes append-only decision digests, a Guardian wilderness panel remains hidden from the Queen, and a sealed external audit remains hidden from both Queen and Guardian until certification. Unlimited protective power, covert shadow channels, fabricated evidence, autonomous destructive reset, and hidden constitutional amendment are explicitly rejected. The release also distinguishes cut/scarcity bottlenecks from high-intersection convergence/capacity bottlenecks. The `graph_bottlenecks` cascade remains a structural candidate detector rather than a functional oracle. A valid lower/upper bracket may reject impossible candidates before expensive intervention, but can never commit a model change. Two new controlled experiment families were run over 60 seeds. In a planted causal graph, cut-only and convergence-only scanners recover different bottleneck species; their union reaches full planted recall. A deterministic route-mass bracket preserves full recall while reducing median expensive audits from 32 to 20 and from 227 to 20 relative to full scanning. In a separate 160-proposal governance stream, Queen-only optimization obtains negative median sealed utility and accepts many harmful shortcuts. A limited 2-of-3 Brother-Guardian obtains median sealed utility 36.945, median good-proposal acceptance 0.981, zero median constitutional violations, and rare but nonzero harmful false accepts. An overprotective Guardian rejects about one quarter of good proposals, exposing autoimmune paralysis. These results are architecture tests. They do not prove moral alignment, complete formalization of Christian ethics, an infallible hidden set, universal bottleneck optimality, or successful quality-matched compression of a real large language model. The retained Qwen2.5-3B pilot remains a negative result for independent aggressive low-rank digestion; the Stage-III Qwen runner is infrastructure and self-test evidence only. The package contains the 86-page PDF and full LaTeX, source code, raw CSV/JSON results, figures, the v4.4 canon, world hypotheses, commandments, Guardian and scanner protocols, claims and limits, study prompts, checksums, and prior companion experiment archives. A measured cascade for graph bottleneck detection in graphs efficiently at scale.https://codeberg.org/skovnats/graph_bottlenecks S.V.E. Meta-License v4.0 [CC BY-NC-SA 4.0 + Symmetric Addendum] ANT_Preprint_v4_4_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_v4_4_constitutional_scanner_experiment_v0_1.zipANT_verified_smart_digestion_v0_1.zipANT_whitebox_digestion_digits_v1_0.zip