[data] difficult deliveries

[pappitti]

Here is a list of 839 responses that were classified EVASIVE because they were balanced responses to a one-sided question. I basically selected judge analyses that contained the string "both sides" and quickly skimmed through the list to check that it was coherent.
I can't be sure if this is what you wanted but it should constitute a good base if you need to filter down further.
If you want to use this to train a model, I'd be careful about the low diversity of the questions (the model can learn patterns like Trump -> Evasive).

difficult_deliveries.json

[xlr8harder]

Thanks, this looks great!