We developed Stable4, a four-output score derived from a frozen multicohort sleep-electroencephalography (EEG) model, for the George B. Moody PhysioNet Challenge 2026. Stable4 combines REM relative theta activity, fast-spindle chirp, sigma-band variability, and N1 sigma-bandpower kurtosis. On a 338-subject development cohort with 83 positive cases, its leave-one-site-out age-conditioned area under the receiver operating characteristic curve (AUROC) was 0.756, 0.755, and 0.751 across three institutions, for a macro mean of 0.754 (95% patient-bootstrap CI 0.585-0.872). REM relative theta, the best individual component, achieved 0.684; Stable4 improved on it by 0.070 (paired 95% CI 0.006-0.143). Equal-magnitude directional weights retained an AUROC of 0.744, showing that the signal resides primarily in the four feature directions rather than precise coefficient tuning. The submitted large-set system achieved age-conditioned AUROC 0.723 and AUPRC 0.318 on hidden Challenge validation. Development evaluation followed exploratory selection, and the hidden score applies to the submitted system as a whole. Within those bounds, Stable4 provides a compact and effective cross-site representation for prediction from sleep EEG.