Exploratory internal benchmark of workflow-aware RGB-VNIR-LF-NMR sensing for apple moldy-core screening under simulated incomplete sensor availability.

Apple moldy core is a hidden internal disorder that can escape surface inspection and compromise postharvest sorting. This exploratory study evaluates incomplete multimodal sensing as an internal workflow benchmark rather than evidence of deployable screening performance or superiority over mature Vis/NIR technologies. Using 236 complete tri-modal Red Fuji apple samples with binary post-cut reference labels, we assessed RGB, visible-near-infrared (VNIR) multispectral, and low-field nuclear magnetic resonance (LF-NMR) inputs across all seven non-empty sensor subsets under 15 seeded random apple-level 80/20 internal train/validation partitions with simulated missingness. All deep models followed a fixed 50-epoch schedule, and only the epoch-50 model state was evaluated and reported. Unified Any-Modality had the numerically highest mean Average-subset AUC (0.7423 ± 0.0489), whereas the Weak-subset-aware Deep-score Router had the numerically highest mean Worst-subset AUC (0.6108 ± 0.0695) and Tail-subset AUC (0.6426 ± 0.0581). Four conventional optical baselines achieved native-subset AUCs of 0.7471-0.7699, including 0.7668 ± 0.0568 for VNIR plus PLS-DA. A class-weighted cross-entropy sensitivity analysis on three representative models produced no consistent AUC benefit, and every paired Average-subset AUC interval included zero. Fixed-threshold diagnostics showed low sensitivity for several deep models at 0.5, so no operational screening threshold is claimed. The results support an objective-dependent, workflow-aware interpretation of incomplete sensing while remaining conditional on internal validation, simulated missingness, and asymmetric modality encoders.