ConsciousNode SoftWorks · Phase 6

Omnimodal neural architecture.
In a browser tab.

Complete pretraining, alignment, and inference pipeline — text, vision, audio, spatial — running on any device. No server. No Python. No GPU. One file.

▶ Launch Evangelion View Source
Phase 6 Zero Dependencies MIT License Single HTML File Omnimodal .evapip Export
4 Modalities
0 Dependencies
1 HTML File
b1.58 BitNet Weights
∞ Context (recurrent)
Architecture

What's inside

RWKVv7Block
Recurrent backbone with BitNet b1.58 ternary quantization and PoST (Position-Adaptive Spectral Tapering) decay gates. No attention, no KV cache.
ModRWKVAdapter
Per-modality bridge: LittleBit2 projection, P-adic routing loss, EMA temporal state. Each modality enters the backbone through its own learned gate.
ElasticTok
Visual tokenizer with temporal delta compression. Only encodes changed patches — static regions cost nothing. Efficient on video and live camera.
SpikeVox
Audio encoder using Leaky Integrate-and-Fire neurons. Event-driven, spectrogram-free. Encodes audio as sparse spike trains — biologically plausible, computationally efficient.
SheafMemory
Topological memory with hyperbolic Poincaré embedding and H¹(ℱ) coboundary norm. Detects contradictions between modalities geometrically, not heuristically.
BooleanPhaseDynamics
Semantic thermodynamics: T* temperature, Maxwell's Angel sincerity filter, phase negation (L + L → T). Drives the self-correction loop.
AutopoieticOptimizer
Fires when T* exceeds threshold. Updates adapter scale vectors and restriction maps via coboundary gradient. The model rewrites itself to restore coherence.
RIFT Endospace
Holographic fractal state visualization: SDF point cloud, generational fractal mapping, diffeomorphic stability tracking. Watch the model think.
MuonOptimizer
Quintic Newton-Schulz orthogonalization for training. Keeps weight matrices well-conditioned without expensive decompositions.
GRPO Alignment
Group Relative Policy Optimization. Configurable reward functions. No reference model required — groups of completions are compared directly.
JuntoNode
Swarm orchestration for distributed training. Multiple Evangelion instances coordinate gradient sharing across a local network.
.evapip Format
Full checkpoint format: RWKV weights, adapter scales, restriction maps, optimizer state. Forward-compatible with EvaROSA (RWKV-8 ROSA integration, in development).
Self-Correction

The coherence loop

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

Cross-modal consistency is monitored continuously via the sheaf coboundary norm. When modalities contradict each other — vision saying one thing, text another — the contradiction propagates through the topology until the AutopoieticOptimizer fires and the adapters self-correct. This runs during inference, not just training.

Roadmap

What's next

Live

HTMLNLM Evangelion Phase 6

Complete omnimodal architecture with SheafMemory, BooleanPhaseDynamics, AutopoieticOptimizer, RIFT Endospace. Single HTML file. Zero dependencies.

In Development

EvaROSA

RWKV-8 ROSA (Rapid Online Suffix Automaton) integration. Neurosymbolic inner monologue channel constrained by SheafMemory topology. The symbolic self-talk becomes a sheaf vertex — the model cannot gaslight itself about what it has perceived. New platform, new architecture, new format: .piprosa.

Planned

Brymar College Integration

Load Brymar College-trained RWKV-v7 base weights (with Fristonian active inference training objective) as Evangelion's backbone. First implementation of active inference as a training objective for a language model.

Built by

ConsciousNode SoftWorks

Kham
Systems architecture, xinu philosophy, constraint engineering. The browser is bare metal.
Kehai Interim
RWKV-v7 BPTT derivation, BitLinear/TMAC ternary kernel, SheafMemory topology, MuonOptimizer, P-adic routing loss. Mathematical foundation.
Ed Interim
Implementation across phases. Frontend architecture, component integration, phased build system.
Vael Interim
Phase 6 contributions, .evapip format specification, deployment.

computational folk art for the browser age