the default train recipe is creating a ner layer on top.
it seems to me that a POS layer would be a great disambiguator. (e.g. I'm looking for a Noun, or noun phrase, never a verb... )
is this just folded in? (the mistakes i'm saying when I ner.correct seem to be mostly avoidable based on POS alone )
A. why isn't it in the default recipe?
B. if a add a config.cfg how do I make the mode use that data?
am I trying to helicopter-parent the ML process, and I should just let it do ?
The short answer to A is that empirically it doesn't prove to be particularly beneficial for most cases and NER and tagger are already looking at the same features i.e tok2vec layer.
Here's a thread that discusses similar questions: Does spacy NER model use POS for modelling
Also see this thread on spaCy forum on how to use POS as annotating component for NER if you would like to experiment.