A bicycle-share depot can label only 35 bent-wheel photos this month. Leadership still wants a junior mechanic to see a suggested mark and then accept or override it. They do not want a from-scratch model that needs thousands of new labels. Which approach fits?

With only about thirty-five bent-wheel labels this month, the practical path is a human-in-the-loop assistant or a pretrained vision service where a junior mechanic accepts or overrides each suggested mark. That approach fits because pretrained vision can propose labels with little new data, and human review keeps quality under control without waiting for thousands of fresh annotations. Insisting on a from-scratch supervised model that cannot start until thousands of labels exist blocks useful assistance this month and ignores transfer learning and review workflows leadership already described. Replacing the mechanic with an unreviewed score removes the accept-or-override control the depot required and risks silent mistakes on safety-relevant wheel defects. Training a language model on wheel photos confuses modalities: text-centric models are not a substitute for vision labeling, and claiming they "do not need labels" does not make photo classification work without appropriate training signal. Scarce labels plus mandatory human review is a classic HITL or pretrained-service scenario. Exam caveat: scarce labels plus mandatory human review point to HITL or pretrained services, not a greenfield supervised build that waits forever for labels. Operational check: run the pretrained suggester on a pilot batch, log mechanic overrides as new labels, and only then decide whether collecting more labels for fine-tuning is worth the cost.