We present a CodeBERT-based system for detecting AI-generated code in SemEval-2026 Task 13 Subtask A. To address class imbalance and model overconfidence, we apply stratified balanced subsampling, dynamic per-epoch F1-macro threshold optimization, and label-flip bias correction. The model is trained using TPU-accelerated fine-tuning and achieves a validation F1-macro of 0.874 and a private leaderboard F1-macro of 0.53. Ablation studies confirm the effectiveness of our balancing and calibration strategies under distribution shift.