SAINT-G: controlled AI growth through validated neural grafts

Hi everyone,

I’m developing SAINT-G: Scalable Auditable Intelligence through Neural Grafting.

The project explores whether model growth can be made more modular and auditable by training small structured grafts instead of updating the full model.

Current graft form:

Delta W = A Phi B

The base model remains frozen by default. Candidate grafts are trained, validated against the composed model, accepted/rejected by validation gain, and stored as recomposable artifacts.

Current experimental bridge:

• Backbone: DRM Transformer
• Full baseline: DRM 125M/350M
• Grafted path: DRM 5M + SAINT-G grafts
• Metrics: validation loss, perplexity, VRAM, checkpoint size, recomposition drift, retention/regression

The project is still experimental. It does not claim superiority over LoRA/QLoRA; those are required baselines.

The research question is:

Can validated neural grafting provide a practical path toward controlled, reversible, and auditable model growth?

Repo:

github.com

GitHub - gnai-creator/SAINT-G: Toward controlled evolution of artificial...

Toward controlled evolution of artificial intelligence through validated neural grafting.

I’d be especially interested in feedback on:

• baseline design;
• LoRA/QLoRA comparison protocol;
• retention/regression evals;
• graft registry format;
• safety/control evals for modular model growth.