GUARD-SSL: Class-Conditional Reliability Gating and Balanced Pseudo-Label Admission for Tail-Robust Semi-Supervised Learning

Semi-supervised learning can reduce labeling cost, but pseudo-label self-training introduces a second source of supervision whose errors are produced by the model itself. A fixed confidence threshold is simple, yet it treats every class as equally reliable and can allow majority classes to dominate the pseudo-labeled pool. This paper introduces GUARD-SSL, a reliability-gated self-training method that estimates class-specific pseudo-label thresholds from out-of-fold predictions on the labeled seed and then limits each retraining round with a square-root class-prior admission quota. The reliability statistic is a one-sided Wilson lower confidence bound on predicted-class precision; it is used as a conservative selection heuristic rather than claimed as a formal post-selection guarantee. The gate remains anchored to true labeled data so pseudo-label errors cannot directly relax their own admission criterion. We evaluate GUARD-SSL on eight public real datasets and two controlled synthetic datasets using five matched outer seeds, producing 50 dataset-seed blocks. With 20% of each outer training partition labeled, GUARD-SSL attains 0.7459 mean balanced accuracy, 0.7383 macro-F1, 0.6394 worst-class recall, 0.9217 pseudo-label precision, and 0.4270 pseudo-label coverage. Relative to labeled-only training, balanced accuracy improves by 0.73 percentage points and worst-class recall by 2.04 points; the latter remains significant after Holm correction (p=0.0029). Relative to fixed-threshold self-training at 0.95, GUARD-SSL preserves comparable pseudo-label precision while increasing coverage and improves worst-class recall by 1.18 points (Holm-adjusted p=0.0374). Under 10-20% label noise and minority-label scarcity, its advantage is concentrated in worst-class recall rather than universal average accuracy. Ablations show that removing the class quota increases pseudo-label coverage from 42.7% to 70.6% but reduces worst-class recall from 0.6394 to 0.6201. The results support a narrow conclusion: class-aware admission can make self-training more tail-robust, but it does not eliminate confirmation bias or replace stronger representation-learning methods.