I broke my data. I was running a Knuth shuffle manually on the numpy arrays for input and true values. Apparently the familiar mechanism of TMP=A; A=B; B=TMP; doesn’t work as expected with numpy arrays and sub-arrays, and I wound up losing big swaths of data and replacing them with duplicates. I’ve corrected that, but it invalidates the results of my earlier experiments. I’ve updated earlier articles appropriately.