| # Within-source robustness tables (assembled) | |
| ## Architecture comparison (within-source, full Spice_Spectrum test) | |
| | Model | Top-1 | Top-5 | Macro-F1 | | |
| |---|---|---|---| | |
| | SVM (HOG + Color) | 32.49 | -- | 30.69 | | |
| | ResNet-50 (fine-tuned) | 99.36 | 99.95 | 99.36 | | |
| | EfficientNet-B4 (image-only) | 99.59 | 100.00 | 99.59 | | |
| | ViT-Base/16 (fine-tuned) | 99.73 | 99.95 | 99.73 | | |
| | SpiceFusionNet (ours) | 99.68 | 100.00 | 99.68 | | |
| ## Augmentation ablation (within-source) | |
| | Augmentation | Top-1 | Macro-F1 | | |
| |---|---|---| | |
| | a3_no_aug | 98.55 | 98.54 | | |
| | a3_std_aug | 99.05 | 99.05 | | |
| | a3_spice_aug | 98.91 | 98.91 | | |
| > Note: these are WITHIN-source metrics (they justify the model choice and show the task is non-trivial). The cross-source collapse is reported separately in the shortcut matrix — augmentation does not close it. |