As part of the George B. Moody PhysioNet Challenge 2026, our team, insomniac111, developed two open-source machine-learning pipelines to predict future cognitive impairment from polysomnography (PSG). The public data included a 1,103-record small training set and a 6,600-record large set from three institutions, with heterogeneous physiological signals, Complete AI Sleep Report (CAISR) annotations, human annotations, demographics, and diagnostic labels. After channel harmonization and 200-Hz resampling of EEG, EOG, and ECG, XGBoost used a compact 156-feature representation: 12 demographic variables, 105 physiological summaries, 31 CAISR measures, and eight missingness indicators. TabPFN V2 used an expanded, quantile-normalized table of 243 features that additionally included regional EEG spectra, heart-rate variability, enhanced oxygenation, human annotations, and stage-conditional physiology. We used leave-one-site-out validation to estimate cross-source generalization. On the blinded evaluation, small-set trained XGBoost achieved an age-conditioned AUROC of 0.765, large-set XGBoost achieved 0.716, and small-set TabPFN V2 achieved 0.645. Notably, large-set XGBoost slightly improved conventional AUROC over small-set XGBoost (0.855 versus 0.851) while reducing the official score. This divergence suggests that additional data strengthened global discrimination but not ranking among age-matched patients. Possible causes include greater reliance on age or site-associated signals, training-to-evaluation shift, and hyperparameter or objective mismatch. These findings favor metric-aligned sampling, feature-matched ablations, and site-aware blending over an assumption that increasing training-set size alone will improve Challenge performance.