The model had been trained on text from a decade in which certain outcomes were considered inevitable. Its predictions were accurate and confident. The outcomes it had been trained on had not occurred. No one had retrained it. It was still predicting.

The fence had been erected in 1962 along the property line between the Pruitt parcel and the Delaney parcel, and in sixty-three years it had been repaired seven times by four different people, two of whom were alive. The current Pruitts and the current Delaneys did not speak to each other. The original cause of the dispute had been a well, and the well had been filled in during a drought in 1991, but the dispute had long since transferred itself to the fence. What the fence had observed, if a fence could observe: both families used the same corner of their respective yards, the corner nearest the fence, more than any other part of their properties. The children on both sides had made games involving the fence as a boundary marker. The teenagers on both sides had leaned against it at different times in the same years. One Delaney daughter and one Pruitt son had spoken to each other through it for most of one summer before they stopped, for reasons the fence did not know. The fence was not in poor condition. It had been maintained. It was not clear to a neutral observer what function it served, since both families had larger yards on the far side of their respective houses, away from the property line, which they used far less. The fence, if asked, would have had no answer for this. It was not in the business of explanations.