Why automated drilling needs close pairs
A simplified version of the paper's simulation, drawn to be looked at rather than read. Holes go into a made-up ore field four at a time, wherever they cut the model's uncertainty most, so they spread out. The blue rings show how far the model thinks grade carries from each hole. With every hole far from every other there's nothing close enough to compare, and the rings stay blurred. Each field gets drilled twice, without the guard and then with it: the guard adds a couple of close pairs, and the rings settle.
In the paper's simulations from four corner holes, about half of first batches left that distance undetermined, and the usual prediction scores didn't notice. The guard fixes it early, with no accuracy cost we could detect. The details are in the paper.