⚠️ ARCHIVED — Historical / Experimental Project

This repository is kept for historical and educational purposes only.

What was wrong

This project was built on the incorrect assumption that an LLM's reasoning engine and its knowledge structure could be cleanly separated and independently trained as a database. This turned out to be a fundamental misunderstanding of how knowledge and reasoning are entangled in transformer architectures.

What I learned

The failure of this approach directly led to the development of Verantyx — an LLM-free symbolic reasoning engine that takes a completely different path: instead of trying to restructure LLM internals, it builds reasoning from first principles using Cross-structure grammars, atom-based fact extraction, and formal verification.

Current work

• Verantyx V6: github.com/Ag3497120/verantyx-v6 — LLM-free symbolic reasoning engine, scoring on HLE benchmark
• HLE Results: huggingface.co/datasets/kofdai/verantyx-hle-2.6

Do not use this repository as a reference for current architecture. The ideas here were an important step in the exploration, but the approach was wrong. See Verantyx for the updated direction.

NullAI: Revolutionary Multi-Domain Knowledge System

🌐 Live Applications

Dendritic Memory Editor - Web Application

• Create and edit .iath knowledge tiles directly in your browser
• Interactive 3D coordinate visualizer
• No installation required - works on any device
• Perfect for domain experts, researchers, and educators

Base Model: Microsoft Phi-4 14B

🌟 Key Innovations

1. Knowledge Tile System (知識タイル)

• Each tile represents a specific piece of knowledge with clear boundaries
• Tiles are independently verifiable and traceable to source
• Can be updated, validated, or removed without retraining the model
• Enables transparent knowledge provenance

2. Spatial Knowledge Encoding (空間座標記憶)

• X-axis: Specificity (general → specialized)
• Y-axis: Certainty (uncertain → verified)
• Z-axis: Domain (medical, legal, science, etc.)
• Additional dimensions: Temporal relevance, source credibility, complexity level
• Semantic relationships automatically emerge through spatial proximity
• Enables intuitive navigation through knowledge space

3. Judge System - Alpha & Beta Lobes (判定システム)

• Verifies factual consistency
• Cross-references with knowledge tile database
• Checks logical coherence
• Validates causal relationships

• Identifies contradictions
• Detects fabricated information
• Flags uncertain claims
• Monitors confidence boundaries

4. ORCID-Based Expert Authentication

• Knowledge tiles validated by domain experts
• Experts authenticated via ORCID (Open Researcher and Contributor ID)
• Verification status tracked and displayed:
  • 🟢 Expert Verified
  • 🔵 Community Reviewed
  • ⚪ Unverified
• Continuous expert review and updates

5. Zero-Hallucination Architecture

1. Retrieval-based (not generative) knowledge sourcing
2. Expert verification before tile creation
3. Real-time Judge System validation
4. Confidence scoring for uncertainty quantification
5. Transparent reasoning chain display

6. Rapid Specialized AI Creation

• Select target domain (medical, legal, education, etc.)
• System automatically configures:
  • Relevant knowledge tile subset
  • Domain-specific validation rules
  • Expert verification pipeline
  • Specialized prompt engineering
• No model retraining required
• Instant deployment capability

7. Transparent Confidence Scoring

• Overall confidence percentage
• Contributing tile confidence scores
• Hallucination risk assessment
• Knowledge coverage metrics
• Expert verification status

8. Episodic Binding & Context Management

• Layer 2 Episodic Binding for conversation continuity
• Layer 5 State Management for long-term interaction
• Context-aware tile retrieval
• Memory consolidation across sessions

🏗️ Technical Architecture

Core Components

┌─────────────────────────────────────────────────────────┐
│                    User Interface                        │
│              (Web / API / CLI / HuggingFace)            │
└─────────────────────┬───────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────────────────┐
│              Inference Engine (Runner)                   │
│  • Query Processing • Tile Retrieval • Response Synthesis│
└─────────────────────┬───────────────────────────────────┘
                      │
        ┌─────────────┴─────────────┐
        ▼                           ▼
┌──────────────────┐      ┌──────────────────┐
│  Judge System    │      │ Knowledge Tiles   │
│                  │      │   Database        │
│  Alpha Lobe  ✓   │◄────►│                  │
│  Beta Lobe   ✓   │      │  • 16K+ tiles    │
│                  │      │  • Spatial index │
│  Validation      │      │  • ORCID links   │
└──────────────────┘      └──────────────────┘
        │                           │
        ▼                           ▼
┌──────────────────────────────────────────┐
│     Base Model: DeepSeek R1 32B          │
│     (Used for understanding & synthesis)  │
└──────────────────────────────────────────┘

Data Flow

1. Query Input → User asks a question in natural language
2. Intent Analysis → System determines domain and knowledge requirements
3. Tile Retrieval → Relevant tiles fetched from multi-dimensional space
4. Alpha Lobe Check → Logical consistency validation
5. Synthesis → DeepSeek R1 combines tiles into coherent response
6. Beta Lobe Check → Hallucination detection scan
7. Confidence Scoring → Uncertainty quantification
8. Response Output → Answer with full metadata and transparency

📊 Specifications

Model Information

• Base Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
• Parameters: 32 billion
• Quantization: 8-bit (optional, for resource-constrained deployment)
• Context Window: 32K tokens
• Languages: Primary English, with multilingual tile support

System Requirements

• Minimum RAM: 64GB (for 32B model)
• Recommended RAM: 128GB
• Storage: 100GB+ (model + knowledge base)
• GPU: NVIDIA A100/H100 recommended (CPU inference supported but slower)

Knowledge Base

• Total Tiles: 16,503+ (continuously growing)
• Domains: 55+ specialized areas
• Expert Contributors: 342+ ORCID-verified experts
• Average Confidence: 87.3%
• Update Frequency: Real-time as new tiles are verified

Supported Domains

🎯 Use Cases

1. Educational AI Tutors

• Deploy subject-specific tutors in minutes
• Expert-verified educational content
• Adaptive learning with confidence feedback
• Safe for K-12 and higher education

2. Medical Information Systems

• Clinical decision support with expert validation
• Evidence-based medical knowledge
• Always recommends professional consultation
• Tracks source citations and confidence

3. Legal Research Assistants

• Case law and statute retrieval
• Multi-jurisdiction support
• Expert attorney validation
• Clear disclaimers and limitations

4. Enterprise Knowledge Management

• Internal knowledge base integration
• Expert-verified company information
• Secure deployment options
• Custom domain specialization

5. Research & Development

• Literature review assistance
• Cross-domain knowledge synthesis
• Citation tracking and verification
• Collaboration with subject matter experts

📈 Performance Metrics

⚠️ Limitations & Disclaimers

Current Limitations

• Knowledge base coverage varies by domain
• Expert verification process introduces latency for new information
• System performance depends on tile quality and coverage
• Not a replacement for professional advice in critical domains (medical, legal)

Important Disclaimers

• Medical: Always consult qualified healthcare professionals for medical decisions
• Legal: Not a substitute for licensed legal counsel
• Financial: Not financial advice; consult certified financial advisors
• General: Verify critical information through multiple sources

📄 License

Base Model License

• Microsoft Phi-4: MIT License
• See microsoft/phi-4

🌐 Live Applications & Resources

Web Applications

• Dendritic Memory Editor (Web-Based): https://dendritic-memory-editor.pages.dev/#/
  • Create and edit .iath knowledge tiles in your browser
  • Interactive 3D coordinate visualizer
  • No installation required
• HuggingFace Spaces Demo: https://huggingface.co/spaces/kofdai/null-ai
  • Full-stack NullAI application
  • Interactive inference interface
  • Knowledge management tools

Code & Documentation

• GitHub Repository: https://github.com/Ag3497120/nullai-phi-4-14b-v2
• HuggingFace Model: https://huggingface.co/kofdai/nullai-phi-4-14b-v2

Contact & Support

• Developer: Kodai Motonishi (@Ag3497120)
• Email: kodai820820@gmail.com
• Issues: GitHub Issues

📊 Citation