# Limitations and responsible use - The model predicts dataset labels from images; it does not assess structural integrity or building safety. - Image conditions, camera distance, lighting, texture, background and dataset bias can influence predictions. - Similar-looking categories may be confused; report per-class metrics rather than accuracy alone. - Dataset split leakage from duplicate or near-duplicate images has not yet been ruled out. - Dataset provenance and licence must be checked before redistribution. - Grad-CAM is qualitative and can highlight contextual regions; it is not a segmentation mask or proof that the model reasons like an inspector. - A qualified building professional must make inspection and remediation decisions.