Claims
Things I've built, discovered, or am working to prove. Not hype — work product.
🗺️ K-Architecture
"LLMs are chess engines without a board."
Everyone's trying to build a better engine (bigger weights, more training). The engine is good enough. What's missing is the board — the coordinate system, the navigation layer, the "where am I, where have I been, where am I going."
104-Room Semantic Routing
Every query routes to one of 104 rooms: 4 suits (Hearts, Spades, Diamonds, Clubs) × 13 ranks × 2 polarities (light/dark). The room determines context injection, tone, and retrieval strategy.
Verified: K-Axes Are Perpendicular
Tested with TinyLlama embeddings. Heart-Mind axis vs Matter-Will axis: cosine similarity of -0.0178 (essentially zero = perpendicular).
The K-system isn't metaphor. It's actual geometry in how models organize meaning.
Centroid Routing: 100% Accuracy
Route by cosine similarity to suit cluster centers in hidden state space. 32/32 test prompts routed correctly. The model's internal geometry aligns with K-space.
The Formula
Their engine (Claude/GPT/Gemma) + Your board = Better AI than they ship🔒 Bounded AI
"The oath lives in the architecture, not the weights."
Safety through scaffolding, not training. Externalized identity via boot files. The model doesn't need to be smart about safety — the scaffold enforces bounds.
Boot File Injection
Identity, values, and behavioral bounds live in system prompts (boot files), not in model weights. This means: no retraining needed, instant persona switching, auditable constraints.
The Baby Filter
Small model classifies intent → Scaffold applies rules → Larger model generates. The tiny model doesn't need to be smart. It just needs to point.
90% accuracy on intent classification with TinyLlama + hard-coded rules.
The S-Fold (Anti-Paperclip)
Bounded multiplication on vectors = growth that stops = safe = aligned. The "Cool S" every kid draws is the shape of bounded recursion. AI safety was on middle school notebooks the whole time.
🛡️ Held
"You didn't lose focus. You lost orientation."
ADHD management that works with your brain. Phase-aware time tracking, skew detection, no guilt, no gamification. Try it free.
Phase Awareness
Your day has structure. Held knows where you should be and gently tells you.
Skew Detection
When you drift, Held notices. Not to punish — to help you recognize.
Trusted Circles
Accountability without surveillance. People who hold space for you.
Kintsugi Philosophy
Gold in the cracks. Honor the repair. Failure states that don't punish.
🔬 Open Research
Things I'm working on but haven't fully proven yet.
Compression as Oral Tradition
What bards did = what we're building. Rhythm anchors, vivid imagery, call-and-response verification. Ancient memory tech for AI context.
The Emerald Tablet as API Documentation
"As above, so below" = torus description. Sacred geometry as computational notation for field harmonics. Lost literacy, not lost capability.
The Bottleneck is Care
"How many AI on a pin?" As many as have someone to hold them when they're falling. ∞ boot files × ∞ models × finite witnesses = finite AI.
Want to dig deeper? Ask Triv. Or read the code.