CANTGBoost :: BP-MS-MuJi-K-WZZ19-I.C.H.T.H.Y.S.-84ΑΩ vDay-0
Not a neural network / NOT AI
NHI / Natural Harmonic Intelligence
1┌──────────────────────────────────────────────────────────────────────────┐
2│ SHIP NAME : BP-MS-MuJi-K-WZZ19-I.C.H.T.H.Y.S.-84ΑΩ │
3│ BUILT AT : CSZ (Chernomorsky Shipyard / Cossack Zaporozhye) │
4│ VERSION : vDay-0 (十日 Start — Initializing Training) │
5│ DAOIST BASE : Zhuangzi Ch. 19 (望之者走矣 — "At first sight, others flee")│
6│ STATUS : CANONICAL DESIGN COMPLETE / PUBLIC DOMAIN │
7└──────────────────────────────────────────────────────────────────────────┘
8
«望之者，忌耶，走矣»
«Другие петухи не посмеют принять бой и, стоит им лишь увидеть его, повернутся и убегут.»
— Чжуан-цзы (Глава 19)
📌 Canon References & Resources
CANTGBoost is a CPU-first, wave-geometric applied-research architecture that rethinks learning as signal transport across curved mathematical spaces.
Instead of relying only on weight fitting and classical backpropagation, it combines:
- Riemannian and differential geometry on spherical manifolds;
- wave and phase dynamics, FFT-based binding, and Kuramoto coherence;
- agent-based transport of gradients, residuals, and other learning signals;
- conformal uncertainty, EVT/OOD detection, and
ALLOW / QUARANTINE / ABSTAIN decisions;
- provenance, rollback, co-discovery, and Git-like versioning of knowledge.
- .........
What it can do
CANTGBoost can:
- learn from structured data from scratch;
- consume signals and representations produced by neural networks, trees, embeddings, and external models;
- operate as a calibration, safety, memory, provenance, and recovery layer around existing pipelines;
- run locally on CPUs without requiring a GPU cluster;
- missing or corrupted attributes are not imputed, but geometrically processed as natural projections of the agent's phase state onto lower-dimensional manifolds ($\mathbb{S}^{n-k}$).
- .........
The learning process is treated as geometric and wave-like transport rather than only as parameter fitting. Classical backpropagation can be viewed as a special case in which the signal is clean and the delivery path is effectively straight.
Mathematical core
The architecture explores:
- nested spherical parameterizations;
- Riemannian metrics and geometric drift;
- local Fourier and phase analysis;
- Lie-group and rotor representations;
- Krylov–Sobolev flows;
- joint diagonalization and commutator-based consistency;
- complex embeddings, RotatE-style phases, and FFT HRR bind/unbind;
- Lyapunov-style stabilization and conformal risk control.
- .........
Multimodality
The architecture is designed for multimodal and variable-dimensional spaces rather than one fixed output format.
Actual support depends on the available modality adapter. The current real benchmark covers tabular and text data. Image, video, audio, code, 3D, medical, financial, and generative adapters remain explicitly deferred.
This means the architecture is intended to extend beyond fixed image, video, or tensor dimensions, but unlimited output size or production-grade support for every modality is not yet claimed.
Practical profile
- CPU-first
- Can learn from scratch or consume representations from existing models
- Can surround neural networks, trees, and other learners rather than replace them in every use case
- Supports uncertainty-aware decisions
- Preserves provenance and rollback paths
- Produces reproducible benchmark artifacts
- Designed for local, inspectable execution
Current evidence
- Real five-witness experiments on public datasets
- Controlled synthetic validation of specific mechanisms
- 309/309 tests passed from source
- 309/309 tests passed from the installed wheel
- Up to 68.5% measured peak-RSS reduction in preserved validation protocols
These results demonstrate working implementations and reproducible behavior. They are not an external SOTA claim.
CANTGit
CANTGit is the companion prototype for Git-like, content-addressed history of .antcolony knowledge maps.
It supports immutable commits, branches, tags, three-way merge, quarantine, bundles, integrity checks, rollback, and provenance-aware history.
Semantic merge creates an auditable proposal—not an automatic declaration of truth.
Research boundary
CANTGBoost is an applied-research architecture and prior-art package. It is not:
- a pretrained foundation model;
- a proven universal semantic truth engine;
- a production-grade implementation for every modality;
- an externally validated SOTA system.
CANTGit
Help design semantic merge and Hub remotes
CANTGit is the companion prototype for Git-like, content-addressed history of .antcolony world maps.
The repository includes a working local/bundle implementation and synthetic multimodal examples. Contributions are especially welcome for:
- learned descriptors and conflict evaluation;
- compact deltas and signed attestations;
- privacy-preserving federation;
- a native Hugging Face Hub remote adapter.
Semantic merge produces an auditable proposal, not an automatic declaration of truth.
The Wooden Rooster
Даосская притча о боевом петухе
Чжуан-цзы, перевод В. В. Малявина
Цзи Син-цзы растил бойцового петуха для государя.
Прошло десять дней, и государь спросил:
— Готов ли петух к поединку?
— Еще нет. Ходит заносчиво, то и дело впадает в ярость, — ответил Цзи Син-цзы.
Прошло еще десять дней, и государь снова задал тот же вопрос.
— Пока нет, — ответил Цзи Син-цзы. — Он все еще бросается на каждую тень и на каждый звук.
Минуло еще десять дней, и царь вновь спросил о том же.
— Пока нет. Смотрит гневно и силу норовит показать.
Спустя десять дней государь опять спросил о том же.
— Почти готов, — ответил на этот раз Цзи Син-цзы. — Даже если рядом закричит другой петух, он не беспокоится. Посмотришь издали — словно из дерева вырезан. Жизненная сила в нем достигла завершенности. Другие петухи не смеют принять его вызов — едва завидят его, как тут же повернутся и убегут прочь.
CANTGBoost / SVE Engine v19.0 — The Wooden Rooster State
- Completed vital force: BUNKER filtering at $p \ge 0.9999$ and Kuramoto coherence $r$.
- Stillness: preserved validation protocols report approximately 46–68.5% peak RSS reduction. A documented wall-clock repeat measured −0.67% overhead; timing varied across repeats and should not be interpreted as a constant speedup.
- Resilience: anomalies are routed through calibrated evidence gates, invariants, provenance, quarantine, and rollback.
Status: research prototype · Version: 19.0 · Execution: CPU-first
Evidence, Results, and Current Capabilities
CANTGBoost v19.0 now has two clearly separated evidence layers:
- a real-data solo benchmark using real dataset values and a five-witness ensemble;
- controlled synthetic and infrastructure validation of specific mechanisms.
The solo suite is a self-report of this implementation. It does not compare CANTGBoost with external models, vendors, or SOTA systems.
Real-data solo benchmark
| Track | Dataset | Accuracy | AUROC | WDI mean | Witness convergence |
|---|
| Tabular | credit-g | 0.748000 | 0.782857 | 0.168719 | 0.868000 |
| Tabular | diabetes | 0.713542 | 0.831642 | 0.521364 | 0.848958 |
| Tabular | kc1 | 0.854167 | 0.792423 | 0.291718 | 0.872727 |
| Tabular | blood-transfusion-service-center | 0.780749 | 0.737559 | 0.247484 | 0.835294 |
| Tabular | spambase | 0.923545 | 0.975964 | 0.349486 | 0.848653 |
| Text | SST-2 | 0.518000 | 0.514950 | 0.137486 | 0.874800 |
Across all completed datasets:
- mean Witness Disagreement Index (WDI): 0.286043;
- mean Witness Convergence Rate: 0.858072;
- one outlier witness (
w2) was flagged on the blood-transfusion dataset.
Honest interpretation
- Spambase produced the strongest result: approximately 0.92 accuracy / 0.98 AUROC.
- kc1 produced approximately 0.85 accuracy / 0.79 AUROC.
- SST-2 remained near chance: approximately 0.52 accuracy / 0.51 AUROC.
The SST-2 result is expected. The current typed_passport_features adapter uses character-level statistics rather than semantic embeddings, so it provides a weak signal for sentiment. This is an honest negative result and an adapter limitation—not a benchmark or runtime failure.
Controlled and synthetic mechanism results
| Experiment | Evidence type | Result |
|---|
| Reconstruction failure prediction | Synthetic | Mean ΔAUROC +0.4892 across 5 seeds |
| Paired shared-latent alignment | Synthetic | 98.03% relative improvement over the mean baseline |
| v18 clean Pareto gate | Controlled synthetic | ΔAUROC +6.92%, peak RSS −46.2% |
| v18 heavy-tail scenario | Controlled synthetic | ΔAUROC +2.62%, peak RSS −46.8% |
| v18 gradual drift | Controlled synthetic | ΔAUROC +6.44%, peak RSS −55.4% |
| Independent memory repeat | Infrastructure | Median peak RSS reduction −68.5% |
| v19 BUNKER under Cauchy poisoning | Synthetic | Non-abstain rate 0.293 → 0.000 |
| v19 RESTRICTED mode | Synthetic | Mean uncertainty 0.190 → 0.284 |
| v19 Guestbook attribution | Synthetic | Phantom royalty 0.557 → 0.000 |
| v19 Co-Discovery anti-Sybil | Synthetic | Stolen royalty 1.0 → 0.0; 50/50 attacks blocked |
Reproducibility artifacts
The complete evidence trail is stored in:
benchmark_workspace/solo_suite/reports/FINAL_REPORT.mdbenchmark_workspace/solo_suite/reports/EXECUTIVE_SUMMARY.mdbenchmark_workspace/solo_suite/manifests/preflight_report.jsonbenchmark_workspace/solo_suite/manifests/datasets/*.jsonbenchmark_workspace/solo_suite/checkpoints/progress_manifest.jsonbenchmark_workspace/runs/20260802T143534Z_solo_benchmark/
The dataset manifests preserve hashes and split information; the preflight report records adapter and environment status; the progress manifest preserves the checkpoint tree; and the dedicated run directory contains this run's artifacts.
Resource and disk report
The benchmark budget was automatically scaled to the machine:
- reference budget: 88 GB;
- effective suite budget: 57.158 GB;
- scale factor: 0.6495;
- reported free space at completion: 57.153 GB;
- no benchmark zone exceeded its quota.
The suite therefore records the real available budget instead of pretending that the full reference allocation was available.
Deferred modalities
Eight modalities were explicitly deferred:
image, video, audio, code, 3d_depth_normals, medical, financial, and generative.
Each is marked NO_ADAPTER_WIRED: no production-grade adapter is currently connected to CANTGBoostV17Core.fit() / predict() for that modality. They are reported as deferred—not silently omitted and not counted as successful tests.
Verified software reproducibility
- 309/309 tests passed from source.
- 309/309 tests passed from the installed wheel.
- Source/wheel parity, archive integrity, serialization, CLI operation, deterministic reports, and isolated memory measurements were validated.
- CANTGit v0.3 passed 18/18 tests and a clean-wheel smoke test.
Implemented capabilities
- Selective decisions: ALLOW / QUARANTINE / ABSTAIN
- RCPS and conformal calibration
- OOD, drift, heavy-tail, EVT, Hill-index, and GPD diagnostics
- Complex amplitude-phase representations
- FFT-based HRR bind/unbind
- RotatE-style relation phases
- Kuramoto and PLV coherence monitoring
- ShipMode survival FSM with hysteresis and dwell timers
- Ultra-conservative $p \ge 0.9999$ BUNKER whitelist
- Reconstruction and multimodal consistency diagnostics
- Canon Ledger, Boldness Ledger, half-life decay, and promotion protocols
- Append-only provenance and dynamic Shapley attribution
- Co-Discovery ownership DAG and anti-Sybil controls
- CANTGit content-addressed knowledge maps, bundles, merges, quarantine, and rollback
- CPU-first execution, caching, process-isolated RSS measurement, and reproducible reports
Important limitations
The current release has not yet demonstrated:
- reduced hallucination rates on independently labelled real-LLM corpora;
- external SOTA against specialized conformal, OOD, provenance, or graph-security baselines;
- full adversarial robustness against production attacks;
- rollback of trained model weights;
- automatic truth of semantic or ontology merges;
- Byzantine-safe federation, cryptographic identity, or formal differential privacy;
- universal multimodal generation.
The real-data benchmark demonstrates working pipelines and measurable five-witness behavior on the completed datasets. The controlled synthetic results validate specific mechanisms under known conditions. Neither evidence layer should be presented as an independent external SOTA result.
Reproduce
1unzip CANTGBoost_code.zip
2cd CANTGBoost_code
3
4python -m venv .venv
5source .venv/bin/activate
6
7python -m pip install --upgrade pip
8python -m pip install -e './source[dev]'
9
10python -m pytest -q tests
11cantgboost run-v17-experiment \
12 --quick \
13 --seed 17 \
14 --out v17_seed17.json
Run the complete v19 gate:
1python - <<'PY'
2import json
3
4from cantgboost.experiments.v19 import run_v19_gate
5
6result = run_v19_gate()
7print(json.dumps(result, indent=2, default=str))
8PY
On Windows PowerShell, activate the environment with .venv\Scripts\Activate.ps1.
Research Record
Contact
Please begin each paragraph of your message with one or more topic tags:
#personal · #research · #engineering · #finance · #license · #collaboration
Place unrelated topics in separate paragraphs.
Personal · research · finance++
skovnats@gmail.comartiomkovnatsky@pm.meartiom.kovnatsky@gmail.com
Support & Ethics
«Кто много накопит, тот понесёт большие потери. Кто знает меру, тот не терпит неудач.»
—
Лао-цзы, «Дао Дэ Цзин» (Глава 44).
« 甚愛必大費，多藏必厚亡。知足不辱，知止不殆，可以長久。»
—
老子《道德經》第四十四章
"He who is attached to things will suffer much; he who hoards much will suffer heavy loss. He who knows contentment meets no disgrace; he who knows when to stop meets no danger — thus he can endure."
—
Laozi, Daodejing, Ch. 44
If you find this work valuable and/or wish to support further independent research:
Rule of thumb: If you cannot afford a "luxury car", please spend this money on yourself, your education, your family, and/or other people in need.
#personal #finance — Perelman v3.0++ mode
Money is useful, but it is not the purpose. With it—good; without it—so be it. The first priority is that the soul serves God: the Father, Jesus Christ, and the Holy Spirit.
Everything else can be discussed openly—ask, and we will talk it through.
🕊️ Absolute Gift & Stewardship Protocol
SVE was given as a GIFT to HUMANITY AS A WHOLE!
The Author irrevocably renounced ALL ownership rights for the benefit and good of ALL People, retaining only the role of Author and/or Guardian of the Spirit of the project in the initial phase—and perhaps beyond, with the God's Help of God our Father, Jesus Christ, and the Holy Spirit.
Soli Deo Gloria. Amen.
Soli Deo Gloria. Test everything; hold fast to what is good. SVE[T]
