OBJECTIVE: Brain-computer interfaces (BCIs) based on selective auditory attention aim to restore communication by decoding selective attention from auditory evoked potentials. Clinical translation of such BCIs requires maintaining sufficient decoding performance despite brain-state non-stationarity. Approach: We compared classifiers across four evaluation settings: offline baseline classifiers using shuffled 5-fold cross-validation; a causal classifier using a chronological 20%/80% calibration/test split; simulated real-time deployment with a static classifier calibrated on 20 trials; and simulated real-time deployment with an adaptive recursive least squares (RLS) classifier, evaluated within-subject and in a leave-one-subject-out (LOSO) setting. The analysis used 62-channel electroencephalography recorded from 25 healthy adults (18 retained after artifact rejection). Main results: The best offline baseline classifier, logistic regression with point-to-point features, achieved a mean ROC AUC of 0.75 and an estimated information transfer rate (ITR) of 2.46 bits/min, derived from ROC AUC via a conservative heuristic. Under causal application, performance decreased to ROC AUC = 0.63 and ITR = 0.68 bits/min. In simulated real-time deployment, static classification dropped further to ROC AUC = 0.51, whereas adaptive RLS improved ROC AUC to 0.68 and ITR from 0.14 bits/min to 1.42 bits/min (p < 0.001, Cohen's d > 1.49). In the LOSO setting, RLS achieved ROC AUC = 0.57 and ITR = 0.86 bits/min. The LOSO result further suggests that zero-calibration deployment is feasible, with personalization occurring trial-by-trial. Significance: Brain-state non-stationarity is a major driver of performance decline in auditory BCIs. Lightweight adaptive recalibration substantially restores real-time performance and supports the translational potential of ERP-based communication paradigms.