ANT (Angular-Norm Transport) & Agentized Neural Computation: A Living Colony for Turning Data, Graphs, Models, and Meaning Trajectories into Verified Executable Knowledge | [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. A gradient is not a number, but an entity. Memory is not a table, but a trace. Learning is not fitting, but digestion and search. Градиент — это не число, а сущность. Память — это не таблица, а след. Обучение — это не подгонка, а пищеварение и поиск. ANT — это не ещё одна нейросеть, а колония, которая может «есть» саму «пищу» — данные, сигналы, графы и наблюдения — и превращать найденные в них закономерности в короткие исполняемые законы; уже обученная модель является лишь частным, предварительно «пережёванным» источником, который можно тоже «съесть». В отличие от современных подходов, которые сначала строят огромный суррогат, а затем сжимают его веса, ANT стремится извлекать, проверять и исполнять структуру реальности напрямую, сохраняя только то, что выдержало причинную, долгосрочную и нравственную проверку (аксиомы/максимы колонии). «Люби Истину и ближнего. Предлагай свободно. Проверяй независимо. Не причиняй скрытого вреда. Не давай никому непроверяемой власти. Всё испытывайте; хорошего держитесь.» ANT v4.5 is the integrated final-candidate release of Angular-Norm Transport. It retains the mathematical angular/log-norm transport framework, one-reality residual sharing, manifold and graph transport, sparse trail programs, Queen codebooks, Ariadne whole-object model digestion, dual bottleneck scanning, admissible LP/bracket filtering, named-shock graph repair, `.antcolony` serialization, quality-matched inference, living-colony search, typed world hypotheses, the six-commandment v4.4 constitution, a limited Brother-Guardian, Witness logging, private Queen gate, Guardian wilderness, and sealed external audit. The v4.5 extension introduces protected dissent and horizon-aware verification. The proposed initial colony prior becomes `87 / 10 / 2 / 1`: validated exploitation, local exploration, creator-rebels, and Invariant Auditors. The one-percent figure is a role-share prior rather than an audit-coverage limit. Invariant Auditors receive no reward for compression, latency, Queen approval, popularity, or blocking volume. They search for counterexamples to a seven-commandment operational constitution, including delayed subgroup harm, ecological externalities, provenance loss, access asymmetry, resource runaway, rare triggers, and higher-order feedback effects. A signed auditor signal has a protected direct path to the Brother-Guardian. One credible signal pauses a transaction and freezes evidence, but cannot create an irreversible veto. Quarantine requires independent reproduction, a hard static violation, or a declared quorum. Every signal, timeout, override, and appeal is written to an append-only Witness hash-chain. The Seventh Commandment is “Do not harm even an ant”: people, animals, living systems, and ecosystems may not be treated as disposable fuel, obstacles, or experimental material. It is operationalized through declared stakeholders, finite consequence horizons, uncertainty, reversibility, least-harm fallback, and human review for external actuation. The release explicitly does not claim a computable guarantee of zero harm over an infinite future. The religious terms “holy fool,” “martyr,” “death,” and “resurrection” are retained only as metaphors. Executable objects are non-sentient disposable sandbox workers. A risky proposal is tested in a bounded copy-on-write environment without production credentials; the worker is terminated and returns a signed Resurrection Packet containing before/after hashes, causal-difference hash, environment, seed, tested horizons, suspected commandments, uncertainty, and replay data. No living being, production system, or irreplaceable data is sacrificed. Two new controlled experiment families use 60 deterministic seeds. In a 220-proposal planted stream, Queen-only median harmful-commit rate is 0.610. A targeted invariant-auditor policy using three independent sandbox panels reduces the median harmful-commit rate to 0.010, delayed-harm commits to zero, and median oracle regret to 3.49 while retaining all median good and novel-good proposals. It uses 459 panel calls rather than 657 for a full three-panel audit. An over-absolute negative control accepts no proposal, demonstrating that unbounded precaution without evidentiary proportionality produces complete adaptive paralysis. In a second 120-proposal sandbox stream, in-place probing produces 11.5 median active-state contamination events. Isolated probing produces zero median contamination. Signed Resurrection Packets obtain zero contamination, median harmful commits 1.0, median good acceptance 0.908, and median evidence completeness 0.983. Over-quarantine removes almost all harmful commits but reduces good acceptance to 0.586. These results are synthetic architecture tests. They do not prove moral alignment, Christian theological completeness, software sentience, reliable detection of deceptive alignment, infinite-horizon safety, universal optimality of `87/10/2/1`, 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 remains executable infrastructure and synthetic self-test evidence. The package contains the full PDF and LaTeX, source code, raw CSV/JSON results, figures, protocols, prompts, claims matrix, checksums, and all retained 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_5_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_v4_5_invariant_auditor_experiment_v0_1.zipANT_verified_smart_digestion_v0_1.zipANT_whitebox_digestion_digits_v1_0.zip Главная поправка: настоящий Минотавр в ANT v5.0 В ANT v5.0 окончательно исправлено значение метафоры: Минотавр — не цель, не аттрактор и не завершённый смысл. Это монстр потери смысла: гладкий, уверенный и правдоподобный бред, продолжающий двигаться после того, как причинная или доказательная Нить уже оборвалась. Выход из лабиринта — отдельный объект: конечная траектория, которая выдержала объявленные функциональные, доказательные, причинные, редкие и конституционные проверки. «ANT выпускает муравьёв в разветвляющийся лабиринт следующих слов: они разматывают Нить Ариадны, возвращаются от тупиков и не дают гладкому правдоподобному бреду-Минотавру выдать себя за найденный смысл.»