Dementia affects an estimated 57 million adults world- wide and is forecast to reach 153 million by 2050, yet roughly three in five community cases are never diag- nosed. Sleep offers an early window onto this process, because slow-wave activity, spindle density and the con- tinuity of the night all covary with memory consolidation and with dementia risk. Overnight polysomnography is already recorded at scale in the age band where impair- ment accumulates, but is read for respiratory indications alone. The George B. Moody PhysioNet Challenge 2026 asks whether the same study can also be read for cogni- tion. Its two metrics make opposite demands on age: the age-conditioned AUROC compares only patients of similar age, so a score that merely tracks age is worthless, while the prevalence-weighted reward needs a posterior that has seen age. We satisfy both from one gain-invariant feature matrix using two heads, one age-residualised and one age- retaining, and reached a conditioned AUROC of 0.597 on the hidden validation site.