It sounds like we did not have enough training data for these two label types to get the performance that we wanted.
But at least for No Sidewalk labels, the correctness of a label is usually quite clear cut; I imagine that we could get acceptable performance from the AI validator on No Sidewalk labels given enough training data.
We avoid showing No Sidewalk labels to validators in Project Sidewalk for the most part because it's tedious/repetitive, and most of the No Sidewalk labels are correct anyway. I think that it would be pretty fast to rapidly go through and validate a ton of No Sidewalk labels to get ourselves more training data.