P R O T O E T H / K

Building AI that thinks deeply and acts responsibly.

---

N.O.V.A.

Native Object Vector Architecture

Technical Report: On-Device Agile Reasoning via WebGPU-Accelerated Hebbian Fast Weights

---

1. Abstract

N.O.V.A. (Native Object Vector Architecture) is a browser-resident, fast-weight dialogue stack executing entirely on the WebGPU compute substrate. Unlike static Transformer models, N.O.V.A. performs token-level Hebbian adaptation in real-time, utilizing a symplectic manifold for context retention and "Origami-style" symbolic scoping for structural logic. This architecture eliminates the need for server-side inference, offering a fully transparent, inspection-ready reference implementation for client-side agile reasoning. This report details the system's mathematical formulation, memory dynamics, and deployment methodology.

2. Architectural Specification

N.O.V.A. is engineered to provide a reproducible, modifiable reference for edge-based learning, moving beyond the paradigm of "black-box" turnkey assistants.

2.1 Compute Substrate (src/core/)

• Device.js: A low-level wrapper around the WebGPU Adapter, Device, and Queue.
• NovaTensor: Manages raw GPU buffers with specialized initialization kernels (Xavier, Residual, Unit-Spinor). It implements safe memory lifecycles (read/dispose), ensuring all mathematical operations remain strictly in-browser with zero server dependency.

2.2 The Fast-Weight Cell (src/layers/FastWeight.js)

The core reasoning unit iterates per token, forming a "Thinking Step" loop:

1. Projection: Input embeddings pass through RMSNorm and project into rotational components $(W_r, W_v, W_g)$.
2. Symplectic Flow: Updates a complex-valued manifold using energy-preserving rotations, preventing gradient decay.
3. Gated FFN: A Swish-gated Feed-Forward Network ($W_{up1}, W_{up2}, W_{down}$) processes the manifold state.
4. Active Inference: The system predicts its own next state via $W_{predict}$ and adjusts the manifold based on prediction error.

2.3 Memory & Biasing Dynamics

• Hebbian Consolidation: Fast weights capture transient token-to-token associations dynamically, governed by variable decay and learning rates.
• HormoneSystem: Tiles affective scalar values across the bias vector, allowing for "emotional" modulation of generation probability.
• OrigamiMemory: A symbolic stack that pushes/pops hidden states upon detecting structural delimiters (e.g., {, }), enabling robust handling of nested logic in code generation.

2.4 Tokenization & Runtime Heuristics

• Entropic Compression: A byte-level BPE compressor learns a compact ~16k vocabulary "gene" table from the corpus, optimizing for high-density concepts.
• Sampling Strategy: Inference is stabilized via src/main.js using n-gram anchoring (bigram/trigram), repetition penalties, anchor bonuses, and mode-specific tags (Mode: chat/task/code).

---

3. Mathematical Formulation

Let the tokenizer map a text stream into tokens $t_1,\dots,t_T$, producing embeddings $x_t = E(t_t)$. The normalized stream is defined as $\hat{h}_t = \mathrm{RMSNorm}(x_t)$.

3.1 Projection & Gating

The fast-weight cell computes the rotational ($r$), value ($v$), and gate ($g$) vectors: $$ r_t = W_r \hat{h}_t,\quad v_t = W_v \hat{h}_t,\quad g_t = \sigma(W_g \hat{h}_t) $$

3.2 Symplectic Memory Flow

State retention is governed by a complex rotation $R(\cdot)$ on the manifold $s$, ensuring long-term stability: $$ s_t = \lambda,R(r_t),s_{t-1} + \beta,[v_t,,\kappa v_t] $$ $$ m_t = g_t \odot \mathrm{read}(s_t,\phi) $$ Where $\lambda$ represents manifold decay and $\phi$ is the readout mixture coefficient.

3.3 Gated FFN & Active Inference

The manifold output is processed via a Swish-gated block: $$ u_t = \mathrm{swish}(W_{u1} m_t) \odot (W_{u2} m_t),\quad c_t = W_d u_t $$ $$ h_t = \mathrm{RMSNorm}(m_t + c_t) $$ The system performs Active Inference, adjusting the manifold towards predicted embeddings $\hat{x}_{t+1} = W_p h_t$ via gradient-free correction.

3.4 Hebbian Update Rule

The memory manifold and projections are updated in real-time during the forward pass: $$ M \leftarrow \rho M + \eta, (h_t \otimes e_{t+1}) $$ Where $\rho$ is the decay factor and $\eta$ is the learning rate.

3.5 Logit Generation

Final logits are computed by projecting back to the vocabulary space, modulated by n-gram priors and bias vectors: $$ \ell_t = E^\top h_t + b_{\text{ngram}} + b_{\text{bias}} $$

---

4. Deployment & Reproduction

Prerequisites

• Hardware: WebGPU-capable GPU (Integrated or Discrete).
• Software: Recent Chromium-based browser (Chrome/Edge).

Operational Steps

1. Serve Static Files: No backend logic is required. Serve the root directory via any HTTP server.

   # Node.js
   npx http-server .
   # Python
   python -m http.server 8000

2. Initialize Runtime: Navigate to http://localhost:8000 and click "Pulse" to initialize the compute shaders.
3. Training & Inference:
   • Auto-Load: If nova.config.js enables snapshot.autoLoad, the system restores model.snapshot.
   • In-Browser Training: Otherwise, scripts/train_browser.js executes token-by-token Hebbian updates on data/training_data.txt.
4. Configuration: Adjust hyperparameters (dModel, layers, learningRate) in nova.config.js.

---

5. Limitations & Scope

• Experimental Nature: N.O.V.A. is a research prototype focusing on architectural novelty (Fast Weights/WebGPU) rather than scale.
• Data Sensitivity: The model is susceptible to dataset bias and assumes ASCII-formatted chat data.
• Persistence: Model state is transient unless explicitly exported via Snapshots.
• Safety: Minimal guardrails are implemented. Not suitable for safety-critical applications without further review.

---

6. License

Licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). Copyright © 2026 Protoethik Co., Ltd. Use, modification, and networked deployment must comply with AGPL-3.0, including source disclosure requirements.

---

7. Citation

If you use this work in your research, please cite:

Wu, Y. (2026). Context Is Geometry (1.0). Zenodo. https://doi.org/10.5281/zenodo.18366793

Direct link: https://doi.org/10.5281/zenodo.18366793