Visual biases in deep learning models, such as focusing on packaging trays instead of meat texture, reduce the reliability of computer vision systems in food safety applications. This study proposes a Grad-CAM-guided bias mitigation framework for multiclass meat freshness classification that combines explainable AI with a lightweight hybrid ensemble design. A MiniCAM attention module is integrated into MobileNetV2 to redirect model focus toward meat-specific visual cues, and its features are fused with complementary embeddings extracted from Xception. The final decision is obtained by combining the predictions of MobileNetV2 with classical classifiers (SVM and XGBoost) using test-time augmentation and grid-optimized weighted ensembling. The proposed framework achieves 99.78% accuracy on the held-out test set and 99.66% ± 0.23 average accuracy under 5-fold cross-validation, while maintaining real-time efficiency (4.3M parameters, 16.5 MB model size, and 825.1 FPS on a single GPU), and effectively suppresses non-informative background elements (e.g., packaging trays) as confirmed by Grad-CAM visualizations. These results demonstrate that integrating explainable bias mitigation with lightweight ensemble learning enables reliable and deployable meat freshness assessment for real-world food safety inspection.