[rspeer]

I'm not following all of the theory involved, but this makes me wonder if the "firewall" is just an effect of the scale on which they're measuring positions.

Chess engines evaluate a large range of positions as having value 0.00. In fact, an omniscient chess engine would evaluate every position as either having value 0.00 or as being "mate in N" (where N could be unreasonably large).

So maybe the blunders from positive positions are smaller because, as evaluated by a sufficiently strong chess engine, they mostly drop you from +(small number) to 0.