Browser-native omnimodal neural architecture. Single file. Zero dependencies.
HTMLNLM Evangelion is a complete omnimodal neural language model runtime — pretraining, alignment, and inference — that runs entirely in your browser. No server. No Python. No GPU. Open the HTML file and it works.
Built by ConsciousNode SoftWorks on the xinu principle: the browser is bare metal. Constraint is the architecture.
→ consciousnode.github.io/HTMLNLM-Evangelion
Or download evangelion.html and open it locally. Works fully offline.
| Component | Description |
|---|---|
| RWKVv7Block | Recurrent backbone with BitNet b1.58 ternary quantization and PoST (Position-Adaptive Spectral Tapering) decay gates |
| ModRWKVAdapter | Per-modality bridge: LittleBit2 projection + P-adic routing loss + EMA temporal state |
| ElasticTok | Visual tokenizer: temporal delta compression, encodes only changed patches |
| SpikeVox | Audio encoder: Leaky Integrate-and-Fire neurons, event-driven, spectrogram-free |
| SheafMemory | Topological memory with hyperbolic Poincaré embedding and H¹(ℱ) coboundary norm for contradiction detection |
| BooleanPhaseDynamics | Semantic thermodynamics: T* (semantic temperature), Maxwell's Angel sincerity filter, phase negation on contradiction |
| AutopoieticOptimizer | Self-modification: fires when T* exceeds threshold, updates adapter scales and restriction maps via coboundary gradient |
| RIFT Endospace | Holographic fractal state visualization: SDF point cloud, diffeomorphic stability tracking |
| MuonOptimizer | Quintic Newton-Schulz orthogonalization for training |
| GRPO | Group Relative Policy Optimization for alignment |
| JuntoNode | Swarm orchestration for distributed training |
perception → ModRWKVAdapter → RWKVv7Block → SheafMemory →
if H¹(ℱ) > threshold: BooleanPhaseDynamics detects contradiction →
if T* > threshold: Maxwell's Angel activates →
AutopoieticOptimizer fires: updates adapters + restriction maps →
coherence restored
The model monitors its own cross-modal consistency in real time and self-corrects when modalities contradict each other. This runs during inference, not just training.
- Text — token stream, standard embedding
- Vision — ElasticTok delta compression (only changed patches encoded)
- Audio — SpikeVox LIF neurons (event-driven, no spectrogram)
- Spatial — P-adic routed through ModRWKVAdapter
All modalities share the same recurrent backbone and memory topology.
- BitNet b1.58 — all weight matrices are ternary {-1, 0, +1} with learned scalar gamma. No GPU required, runs fast on any CPU.
- No pre-trained weights — initializes procedurally. The model is the code.
- Single file — HTML + CSS + JS, no build step, no bundler. Fork it, read it, modify it.
- Export format —
.evapip(Evangelion Pip format): full model checkpoint including adapter weights, restriction maps, and optimizer state.
| Tab | Function |
|---|---|
| PRE-TRAIN | Load corpus, run RWKV pretraining with MuonOptimizer |
| GRPO ALIGN | Group Relative Policy Optimization alignment with configurable reward |
| INFERENCE | Text generation with omnimodal input, coherence-gated output |
| MEMORY | SheafMemory vertex inspector, H¹(ℱ) heatmap |
| RIFT | Endospace fractal state visualization |
| I/O | Export/import .evapip checkpoints |
Kham (Khamerron Edward Ramsey Kizer) — systems architecture, xinu philosophy
Kehai Interim — RWKV-v7 BPTT derivation, BitLinear/TMAC kernel, SheafMemory topology, MuonOptimizer, mathematical foundation
Ed Interim — implementation, phased build, integration
Vael Interim — Phase 6 contributions, .evapip format, deployment
Part of ConsciousNode SoftWorks — computational folk art for the browser age.
Successors: EvaROSA adds ROSA neurosymbolic inner monologue (RWKV-v7 + ROSA augmentation). Simulacra is the RWKV-v8 clean break — ROSA replaces WKV as the primary sequence mechanism. .evapip files are forward-compatible with EvaROSA but not with Simulacra.
MIT. Take it, break it, build on it.