Should be a bit faster, I added some parallel processing.

It has a few gates to go through, it does not just dump the answer from the LLM, but makes sure it is not infected by statist bullshit from training sets.

Not perfect, but the gate is needed :).