As part of the George B. Moody PhysioNet Challenge 2026, we (team NeuroUC3M) developed SIREN (Site-Invariant Rank Ensemble), an algorithm to rank patients by their risk of a cognitive-impairment diagnosis within one to six years from a single overnight polysomnogram. A polysomnogram encodes rich information through many channels, but it also encodes montage conventions and institutional hardware biases, making generalisation to unseen institutions difficult. We therefore enforce site invariance by construction using within-recording percentiles, rank normalisation within institutions, and contrastive augmentations to imitate real-world site discrepancies. SIREN is formed by two branches: an ensemble of logistic, gradient-boosting and LambdaRank heads over 1031 polysomnographic descriptors across five feature views, alongside six frozen self-supervised convolutional encoders over raw electroencephalogram read out by a pairwise attention head. Predictions are then combined via rank averaging. A site probe quantifies the invariance: a classifier separating held-out institutions from the training cohort reaches 0.97 AUC on the raw descriptors but only 0.33 after per-site rank normalisation. On the hidden validation institution, SIREN achieved an age-conditioned AUROC of 0.755, placing our team 6th of 94 teams on the official-phase leaderboard. Test set results were not released at the time of writing.