Semantic coordinates for cross-architecture AI coordination
Trionnx is a symbolic protocol that defines semantic attractors — primitives that organize meaning across heterogeneous AI systems. Different architectures (LLMs, diffusion models) converge on shared semantic regions when interpreting the same Trionnx atom, enabling coordination without explicit message passing.
Atom: 🥚+🔥→🍳|⏱️
Llama interprets: "transformation through applied energy"
Mistral interprets: "raw potential becoming realized form"
Gemma interprets: "the irreversible chemistry of change"
SDXL generates: [image of egg cooking]
SD 1.5 generates: [similar image of egg cooking]
All outputs cluster in the same semantic region.
Trionnx atoms function as universal semantic coordinates.
When multiple AI systems interpret the same atom, their outputs are significantly more similar to each other than to outputs for different atoms — regardless of architecture, training data, or modality.
| Phase | Modality | Architectures | Samples | Effect Size | Significance |
|---|---|---|---|---|---|
| 6 | Text | 6 LLMs | 1,314 | d = 0.85 | p < 10⁻²⁶⁸ |
| 7 | Image | 2 Diffusion | 60 | d = 1.52 | p < 10⁻⁴⁶ |
Effect size interpretation:
- d = 0.2: Small effect
- d = 0.5: Medium effect
- d = 0.8: Large effect
- d = 1.52: Very large effect
| Metric | Text Models | Image Models |
|---|---|---|
| Within-atom similarity | 0.72 | 0.82 |
| Between-atom similarity | 0.66 | 0.69 |
| Gap (attraction strength) | +6-7% | +13% |
| Cross-model agreement | 71-74% | 82% |
Atoms are symbolic primitives that mark regions in semantic space:
| Atom | Domain | Description |
|---|---|---|
PHYS:ΔS↑⏳→∞ |
Physics | Entropy increases over time |
😢→😊|❤️→ |
Psychology | Sadness transforms to joy |
🥚+🔥→🍳|⏱️ |
Transformation | Raw becomes realized through energy |
💰×📈=💎|📉→💸 |
Finance | Investment growth and loss |
∀x∃y: x<y→∞ |
Logic | Infinite progression |
Atoms are not encodings that need to be decoded. They are coordinates that tell AI systems where to create from. The atom defines the region; the agent fills in the details.
Current multi-agent AI systems require:
- Shared vocabularies (brittle, doesn't scale)
- Translation layers (lossy, expensive)
- Common architectures (impractical)
Trionnx atoms provide a shared semantic coordinate system:
Agent A (Llama) writes work → tags with atom
Agent B (Mistral) queries semantic region → finds Agent A's work
No explicit communication required
This enables:
- Routing: Direct work to appropriate specialists
- Discovery: Find related work via semantic proximity
- Coordination: Heterogeneous agents collaborate without orchestration
- Audit: Trace work products to source atoms
Trionnx builds on the Platonic Representation Hypothesis (Huh et al., 2024): that neural networks trained on different data and architectures converge toward shared statistical representations of reality.
Our contribution: This convergence can be addressed symbolically. Atoms are handles into the shared geometry that all capable AI systems discover.
The research literature documents:
- 50-83% cross-model alignment on semantic concepts
- Convergent representations across vision and language
- Shared geometric structure in embedding spaces
Trionnx provides a practical protocol for leveraging this convergence.
| Document | Description |
|---|---|
| Theory | Theoretical foundations and literature connections |
| Phase 6 Results | Text-to-text validation across 6 LLM architectures |
| Phase 7 Results | Text-to-image validation across diffusion models |
| Atom Vocabulary | The 20 validated atoms with descriptions |
| Methodology | Experimental design and statistical approach |
Total validated runs: 8
Total samples: 1,374
Architectures tested: 9+ (LLMs + Diffusion)
Modalities: 2 (Text, Image)
Atoms validated: 20
Lowest p-value: 10⁻²⁶⁸
Highest effect size: d = 1.52
Cross-architecture overlap: 71-82%
- Embedding model dependency: Results measured via BGE-Large (text) and CLIP ViT-L/14 (image)
- Atom vocabulary: 20 atoms tested; generalization to larger vocabularies assumed but not proven
- Production validation: Laboratory results; real-world routing accuracy not yet measured
@misc{page2026trionnx,
author = {Page, Michael},
title = {Trionnx Protocol: Semantic Coordinates for Cross-Architecture AI Coordination},
year = {2026},
publisher = {GitHub},
url = {https://github.com/MemoryforgeAILabs/trionnx-protocol}
}- Huh, M., et al. (2024). The Platonic Representation Hypothesis
- Moschella, L., et al. (2023). Relative Representations for Zero-Shot Stitching
- Bansal, Y., et al. (2021). Revisiting Model Stitching
Apache 2.0 — See LICENSE for details.
MemoryForge AI Labs
For inquiries regarding licensing, collaboration, or implementation:
Michael Page
Founder, MemoryForge AI Labs
Halifax, Nova Scotia, Canada
michael@memoryforgeai.com
"We cannot know if your red is my red. But we can agree that 🔴 means red."
"We cannot know if Llama's entropy is Gemma's entropy. But they agree that PHYS:ΔS↑⏳→∞ marks the same semantic territory."