Preprint tests error checks for faster fusion stability screening

A preprint reports a method for checking machine-learning models that predict fusion plasma stability, including a test that screened a scenario scan in seconds instead of hours. The method adds error ranges, flags inputs unlike the training data and sends doubtful cases to physics simulation software. Tests on a public dataset found that error ranges met their intended coverage on familiar inputs but fell short when plasma shape changed. A few dozen simulations of the new shape, or routing flagged cases to physics calculations, restored coverage. The authors report that not every preregistered hypothesis held.