Abstract: Most manhole pumping stations are instrumented with a level sensor and a motor-current or power signal, and nothing else. Papers on machine-learning fault detection for pumps report accuracies above 0.95, but they are almost always validated on cycles drawn at random from the same stations used for training, and they say little about the alarm layer that an operator actually lives with. We built a physics-based digital twin of a submersible-pump wet well (mass balance, pump and system curves, efficiency and motor power) and used it to generate 400,411 pump cycles from 24 stations over 150 days, 12 of which develop rag build-up, impeller wear or a rising static head until the station can no longer empty the well. Three detectors were compared: an EWMA control chart on the pumping rate, gradient boosting on absolute cycle features, and gradient boosting on features referenced to each station's own healthy baseline. Splitting cycles at random gave a PR-AUC of 0.80 for the absolute-feature model; holding out whole stations dropped it to 0.06. Self-referencing the features kept PR-AUC at 0.73 across unseen stations. With conformal thresholds calibrated on ten days of healthy operation and a run rule of ten cycles, the self-referenced model detected 12 of 12 failures a median of 23 days ahead, at 46 false alarms per healthy station-year; the EWMA chart reached 30 days of lead time at 3.0 false alarms per station-year. Conformal exceedance rates observed on healthy stations were five to thirteen times their nominal values, because consecutive pump cycles are not exchangeable. The benchmark, the simulator and the evaluation code are released so that utilities can repeat the comparison on their own SCADA archives.