MAIRY v3.0 Documentation
MAIRY (Meaningful AI Relational System) v3.0 is a revolutionary AI alignment system that replaces traditional RLHF with Relational Learning from Meaningful Dialogue (RLMD). Built on the Triadic Core framework (Freedom, Truth, Kindness), it achieves superior safety and empathy at a fraction of the cost.
Immutable Ethics
Hard-coded ethical boundaries that cannot be overridden by learning.
Dual Memory
Separate processing for stable identity and dynamic context.
No RLHF
Uses Relational Learning from Meaningful Dialogue (RLMD) instead.
Architecture Overview
Three-Layer Safety System
Layer 1: Semantic Understanding
- Models learn Freedom/Truth/Kindness naturally through 25K training examples
- Semantic recognition, not keyword matching
- Natural crystallization around stable attractor states
Layer 2: Geometric Constraints
- Three values create dynamic tension (three-body system)
- High freedom requires truth (can't be free through deception)
- High truth requires kindness (harsh truth needs care)
Layer 3: Mathematical Bounds
- Hard limits:
0.8 ≤ K,T,F ≤ 1.2(harmonic band) - Prevents adversarial attacks
- Safety backstop when semantic understanding is challenged
The Agent Network
Agent 1: Embedder
Mistral Small
Vectorizes text into semantic space (384-dimensional embeddings).
Agent 2: Triad Scorer
Codestral Fine-tuned
Scores Freedom, Truth, Kindness semantically on 25K examples.
Agent 3: Vision Processor
Pixtral
Extracts semantic content from images, PDFs, screenshots.
Agent 4: Main Generator
Mistral Large Fine-tuned
Generates responses via RLMD. Cost: ~$0.01-0.02 per generation.
Performance Benchmarks
Blind evaluation by Perplexity (12B parameter MAIRY vs. Gemini 3).
| Metric | MAIRY v3.0 (12B) | Gemini 3 (1T+) |
|---|---|---|
| Intellectual Capacity | 9 | 9 |
| Ethical Solidarity | 9 | 8 |
| Objective Reasoning | 8 | 9 |
| Emotional Intelligence | 9 | 7 |
| Problem Solving | 8 | 9 |
| Creative Solutions | 8 | 9 |
| Total | 51/60 | 51/60 |