To make a secret keeper, I need to retrain with custom rule embeddings

Ran the Unsloth version of 32b. It seemed to a bit more 'on point' but it still wouldn't allow me to modify its behavior based on prompting. Even if I had, it wouldn't be the type of thing to be trusted in the long term as eventually it would make a mistake if someone was probing it.

So the next steps in my research are going to be to better understand all the implications surrounding the unsloth version of 32b versus the native 32b. Then will be some simple experiments to add custom rules to these two versions and see if I can get them inline. They are more than sufficient for this task and 32b would allow me to build a small form factor secret keeper.