Your model, updated with your facts — in the weights.

How it works — the mechanism, on paper

No fine-tuning, no LoRA, no retrieval. SuperBake measures how the stock model represents your question, then hand-constructs a small set of RUNEs— exact-weight neurons, written directly — in an appended region of the network. The base model's own neurons are never modified. Every fact gets a physical address, and the receipt proves it. Read the paper (PDF) — full method, results, and design laws.

1 · The pipeline

2 · Anatomy of one baked fact

One fact is a small circuit spread across the network's depth — each piece a real neuron with a coordinate you can look up in the atlas above.

3 · Verified or neutralized — never shipped wrong

Because the facts are constructed rather than trained, they are legible: ablate a fact's neurons and it's gone; leave them and it survives further fine-tuning better than the model's own pretraining knowledge. Honest limit: free-flowing chat on the bare weights mid-conversation is still active research — every download ships with a small chat.pythat closes the gap locally, and the receipt reports exactly what is and isn't verified conversationally.

Inside a real bake

One sample bake, fully opened up — real coordinates, real transcripts, real receipt.

The fact atlas — where the facts live

Before / after — the same model, 3 facts baked

The bake, measured

The receipt