The most useful result was the one I threw out

A drilling record carries where the hole is, how deep it went, what the driller hit on the way down, and what it produced. The commercially interesting question is the last one — how much water will this location give — and that is the one the data refuses to answer.

Depth behaves. Neighbouring wells agree with each other, so location carries real information and a model can use it. Yield does not behave. A well's output is not predicted by its neighbours, and no amount of feature work changed that. Establishing the negative took as long as establishing the positive, and it is the more valuable of the two: it says which promises the business must not make.

Same 49,527 wells, same grid, same method — median of every well in a two-kilometre cell. On the left, well depth. On the right, tested yield. Depth resolves into contiguous shapes; yield does not resolve into anything. That contrast is the finding, and it is measurable rather than a matter of taste.

The aquifer-interval fields are unusually well populated — top and bottom depths present on roughly nine in ten records — and had never been modelled. A leave-one-out skill test put local prediction 38.8% better than the global median. That looked like a new signal.

It was not. The global median is the wrong baseline for a variable that sits at about three-quarters of the well depth. Measured against a depth-derived null instead, the apparent skill disappeared: the aquifer interval is roughly 96% a restatement of the well depth already in hand. It was cut rather than shipped, and if the interval is surfaced at all it should be a derived ratio off the existing depth band, not a modelled target.