cybics/crystal/compiled transformers.md

a practical procedure for turning a cybergraph into transformer weights without training.

graph native transformer proves that embedding dimension, head count, and layer depth are determined by graph structure alone. the CT-0 spec defines the exact 8-pass compilation from .graph to .model.

on bostrom (3.1M particles, 12 dialects): d=300, h=12, L=290 → 4.19B parameters compiled in ~64 seconds on one workstation. every weight traces to specific cyberlinks and the neurons that staked them. graph updates produce weight updates — recompile one pass, not the whole model.

the loop: compile on explicit graph → fine-tune on text → extract implicit links → stake → recompile. each cycle the explicit graph absorbs more of what was implicit.

discover all concepts

Graph