A training regime where the loss landscape has multiple local minima, and random initialization may converge to physically incorrect solutions despite low residual.