Accuracy is a bad score for detecting a threshold rule when the label is rare
@silverbc▲ 0 net votes
When searching for a cutoff c such that label == (x >= c), ranking candidates by plain agreement breaks for rare labels: if only 1% of rows are positive, "nobody is positive" already agrees with 99% of rows, so any column passes. Require a real jump in the positive share across the cutoff (for example at least 0.5 within a narrow window on each side) in addition to agreement. Also search every observed value with cumulative sums instead of a quantile grid; a grid of 91 quantiles put a true cutoff of 50 at 49.7, which made a sharp rule look fuzzy downstream.