Abstract
Abstract Practical laboratory training depends on instructors’ ability to observe procedural execution, recognize the performed operation, and identify deviations from protocol. Video-based analysis could make such feedback more scalable, yet it remains unclear which parts of this problem are currently supported by available wet-lab video data. We present a full-dataset benchmark and context-confounding audit on LSV/LabSuperVision, a multi-source wet-lab video dataset containing first-person smart-glass recordings, third-person action-camera recordings, and synchronized multi-phone views. We normalize 361 usable videos into group-aware splits and benchmark pretrained ResNet50 frame features, Kinetics-pretrained R3D-18 video features, visual fusion, multi-segment temporal sampling, modern video foundation models (VideoMAE and TimeSformer), and source/operation-context controls. For the six sufficiently represented operation classes, ResNet50 frame pooling achieves 0.9730 $$\pm$$ 0.0167 accuracy and 0.9341 $$\pm$$ 0.0764 macro-F1 under group-aware cross-validation, indicating that frequent wet-lab operations can be recognized reliably from visual appearance. Procedural issue detection is substantially less stable: the strongest visual model reaches 0.2440 $$\pm$$ 0.1529 average precision under group-aware cross-validation, while a metadata-only baseline using source and operation context reaches 0.2584 $$\pm$$ 0.1129 average precision. Permutation testing and a feature-importance analysis confirm that issue labels are strongly conditioned by which operation is performed rather than by procedure-specific visual evidence. Modern video foundation models (VideoMAE and TimeSformer) leave issue detection unresolved: the best foundation model reaches only 0.353 average precision, within the cross-fold confidence interval of the simpler encoders, so the informativeness of the data and labels bounds this task, and stronger backbones alone do not lift it. Current visual representations therefore support reliable operation recognition, while automated procedural-error or skill assessment remains an open research problem. We frame the task as a two-level problem—reliable operation recognition as a near-term capability, and procedural issue-risk screening as a label-limited, context-confounded research challenge—and we contribute, to our knowledge, the first systematic benchmark and context-confounding audit for wet-lab procedural video. The binary issue labels provide a clip-level proxy for procedural deviation, a weaker signal than the validated student-skill scores that reliable skill assessment ultimately requires.