A reproducible workflow for diagnosing YOLO26-to-NCNN deployment inconsistencies on resource-constrained Android...

Edge devices increasingly require reliable conversion of desktop YOLO26 models into Android runtimes. This article presents a reproducible YOLO26-to-NCNN diagnosis workflow for resource-constrained Android phones, demonstrated with a 10-class SafeHat PPE task. The method links export, Android asset replacement, Java-JNI-C++ inference, logging, five-stage fault localization, and frozen-image regression. It targets silent deployment errors such as high empty-scene scores, fixed 0.5 confidence, missing geometry, and label disorder. Validation across detection, segmentation, pose, classification, and oriented-box assets shows that letterbox, activation, coordinate-semantic, and layout faults can be localized and repaired, with CPU-path evidence from two Android devices. - Reproducible diagnosis workflow for YOLO26-to-NCNN Android deployment inconsistencies. - Five-stage fault localization from Android symptoms to verified repair artifacts. - Validation package with scripts, frozen images, logs, and CPU-device evidence.