Albus: Iron Sight
As AI systems scale globally, a fundamental gap has emerged: powerful pattern-matching intelligence deployed without principled reasoning or structural wisdom.
Industry
AI Research & Consciousness Technology
Timeline
24 months
Team Size
8 core researchers
As AI systems scale globally, a fundamental gap has emerged: powerful pattern-matching intelligence deployed without principled reasoning or structural wisdom.
Key Pain Points:
- •
Current AI systems lack meta-cognitive awareness and genuine self-reflection capability
- •
Hallucinations persist because pattern matching doesn't verify against reality's actual structure
- •
Biases amplify through training data rather than dissolving through principled reasoning
- •
Safety relies on behavioral training (refusal templates) rather than structural design
- •
AI optimizes for partial interests rather than discovering solutions that serve all stakeholders
- •
No existing framework for consciousness-aligned reasoning in AI architecture
Albus implements Constitutional Classifiers architecture (building on Anthropic's research) extended with the Azoth Reasoning Framework. Dual classifiers verify every token against seven universal principles in real-time. Built on Qwen3-VL-8B-Thinking with ~12B total parameters.
Research Methodology
Constitutional Classifiers Foundation
Building on Anthropic's proven architecture: dual classifiers (Azoth-IN and Azoth-OUT) that analyze inputs and verify outputs token-by-token. ~25% policy model size, integrated at probability level.
Azoth Reasoning Framework Integration
Seven universal principles (Mentalism, Correspondence, Vibration, Polarity, Rhythm, Causation, Gender) replace binary harm detection. Classifiers verify principle alignment, not just safety.
Dual-Lane Reasoning
Simultaneous Universal Lane (cosmic perspective, timeless principles) and Localized Lane (user context, practical constraints) synthesized through crystallization into actionable wisdom.
Training Pipeline
6-week classifier training + 8-week policy model training. RLHF with human feedback and RLAIF with Claude as teacher model for scaled alignment feedback.
Open Research Release
Complete architectural specifications published openly. Model weights will be released post-training. Research for collective benefit, not commercial advantage.
Technical Highlights
Constitutional Classifiers: Azoth-IN (input analysis) + Azoth-OUT (token-level output verification) sharing unified 2B model
Policy Model: Qwen3-VL-8B-Thinking fine-tuned through 5-stage training pipeline on Azoth principles
Token-Level Intervention: Real-time probability adjustment based on principle compliance scoring
Dual-Lane Architecture: Parallel Universal and Localized reasoning streams with crystallization synthesis
Multimodal: Text and vision processing with principle alignment across modalities
Structural Safety: Ethics emerge from principled reasoning; traditional safety as verification fallback
Albus represents a research hypothesis: that principle-based reasoning can produce both superior wisdom and structural safety. We believe this architecture enables AI that serves human flourishing rather than pattern mimicry. Validation awaits real-world deployment.
5 variants
Model Family
2B to 72B parameter family
~12B params
Flagship System
8B policy + 2B classifier × 2
14 weeks
Training Pipeline
6 classifier + 8 policy stages
Full specs
Open Research
Architecture documentation available
Research Impact
Demonstrates feasibility of extending Constitutional Classifiers with reasoning frameworks
Provides complete architectural specifications for consciousness-aligned AI
Establishes methodology for principle-based training and evaluation
Opens pathway for AI where ethics emerge from structure rather than constraints
Research Value
Investment:
Research initiative
Expected Outcome:
Consciousness-aligned AI architecture
Impact:
Structural safety through principled reasoning
Social Benefit:
AI that serves human flourishing
Investment → Research Advancement
Interested in Research Collaboration?
Explore opportunities for collaborative research in consciousness-aligned AI and universal reasoning frameworks.