fulvian/engraft-ngram ? reverse-engineered prompt

Build me the ENGRAFT project from this repo so I can write facts into an LLM’s n gram memory without changing the model weights.

I want a working Python codebase that can train and apply a small overlay file next to a GGUF model, then show that the overlay changes answers only when the right n grams appear and that removing it restores the base model exactly. Please make the main workflow easy to run from the command line, include the data and experiment pieces needed for the Quail examples, and make sure the results and charts can be reproduced from the files in the repo. If there are any missing details in the docs, look up the current llama.cpp or GGUF info online if needed.

It would also be great to have a simple simulator or demo path that lets me compare the base model and the overlay question by question, plus a clean README that explains how to use it without getting too technical.