Wouldn't it be easier to train a vLLM on the handwriting style of the historical person in question? An agent graphologist if you will. Surely there is a lot of pattern matching in the way things are written.

Then again, getting this result from a heavily-generalized SOTA model is pretty incredible too.