Paper 10 · May 2026

CECI: Cross-Embedding Compatibility Index for Model Grafting

William Ken Ohara Stewart

HyperTensor Project · Extended version · TeX source

Abstract

Can skills be surgically extracted from one transformer and grafted into another? We introduce CECI, a gauge-aligned manifold projection measuring geometric compatibility between layers. 120 layer pairs measured on SmolLM2-135M. Within-band grafts (adjacent layers, $\Delta L \le 4$) are viable: Grassmann distance $<0.92$, subspace overlap $\ge 15\%$, gauge alignment $+74\%$. Cross-band grafts (Mix→Refine) are infeasible: GD $>0.96$, gauge $\Delta \approx 0$, residual $>100\%$. Seven Danish-named chimeric models published on HuggingFace. Five of seven improve MMLU over baseline. Cross-model grafting confirmed: Qwen2.5-0.5B FFN in SmolLM2-135M body achieves +6pp MMLU.

1. CECI Feasibility Map

Band PairΔLOverlapGDViable?
Mix→Mix0–224.9%0.89
Compress→Compress0–220.1%0.91
Mix→Compress2–415.4%0.92
Mix→Refine8–127.6%0.96

2. 7 Danish Chimeras

ModelMMLUBoolQ
SmolLM2-135M (baseline)62%40%
minElskede68%53%
minFjollede (cross-model)68%47%

Cross-model grafting: minFjollede uses Qwen2.5-0.5B FFN in SmolLM2-135M body — direct evidence GRC basis transfers functional knowledge.

References

  1. Stewart, W.K.O. Universal Geodesic Taxonomy. HyperTensor Paper XI, 2026.
  2. Stewart, W.K.O. Geodesic Projection Pipeline. HyperTensor Paper II, 2026.