A customer return shows a hairline crack the line has seen twice this quarter. Someone opens a few-shot tool, drops three similar photos into the prompt, and the station starts rejecting. Scrap climbs on glare and machining marks that look “close enough.” The rare class is still thin. The adapter learned the lighting of the examples, not the failure mode that escapes.
Active-learning essays warn about easy-case bias in queues. Inter-annotator agreement owns definition stability. Synthetic-data notes own manufactured variety. This brief is few-shot adaptation for rare plant defects—when sample count is the constraint, not SKU novelty.
Few-shot and related low-shot adapters can stand up a detector when collecting hundreds of labels is unrealistic. They do not waive the quality system. Require a written defect definition before the shots are chosen. Prefer examples that span lighting, orientation, and severity—not three near-duplicates from one shift. Hold a frozen evaluation set that includes confirmed escapes and near-miss lookalikes; score false-reject and escape proxies on that set after every adapter update. If the plant cannot find ten honest negatives that confuse humans, it does not yet have a gate—it has a hope.
Promotion rules that work on the floor: adapter version pinned to the example set hash and the SOP revision; HITL required for the rare class until escape budget is met; no silent overwrite of the previous few-shot pack when a fourth photo arrives mid-shift.
Treat a successful demo on three images as a pilot, not a POR. Rare defects stay rare until the week they are not—and then only a disciplined shot set and a held-out escape panel keep the adapter from automating the wrong cousin.