Those blurry scan numbers made this click for me. 15% of the probability on the three legal values, and high still comes out 40% of the time looking exactly like a confident answer. In my extraction evals I'd also split the test set by input quality, since forced picks on bad scans can hide inside a healthy average the same way one bad schema does.