Target conditioning reveals a skill trade-off in rapid-intensification prediction for western North Pacific typhoons

Rapid intensification (RI) is rare enough that objectives optimized over all tropical cyclone cases may attenuate its wind-change signal. We test whether target conditioning improves RI prediction without degrading population performance and whether learned neural routing contributes skill beyond that conditioning. The retrospective western North Pacific benchmark contains 54 885 IBTrACS cases from 1979–2023 with storm history and monthly ERA5 fields. Models are trained through 2016, selected during 2017–2020, and evaluated independently during 2021–2023. Doubling the RI sample weight reduces the 24-h RI-conditioned wind-change mean absolute error by 4.04 kt, with a 95% storm-block interval of 3.18–4.93 kt, while the full-sample change remains unresolved. A sixfold total weight is the first tested setting with a resolved population penalty. Across 10 random seeds, the learned-gate ensemble yields an average precision of 0.505, a critical success index of 0.320, a Brier score of 0.106, and 24-h full-sample and RI-conditioned errors of 10.94 and 18.83 kt. Every contrast with uniform routing and a single expert includes zero. A history-only learned gate is likewise indistinguishable from one supplied with monthly environmental fields. Target conditioning therefore exposes a measurable conditional–population skill trade-off, whereas learned routing and retrospective monthly fields provide no resolved additional benefit.