A real photo · a real transcription · one real lie
A vision model read this plate.
One number is wrong. Find it.
This is an actual motor nameplate from our test corpus, photographed with glare, and the actual field transcription a local vision model produced — confidently. Exactly one value below doesn't match the plate. Click the one you think it is.
vlm read: 6397-ZZ (confident)
ocr read: 6397-ZZ (the glare fools it too)
human read: 6397-ZZ (our own ground-truther, off this photo)
catalog: 6307-ZZ exists · 6397-ZZ does not
verdict: confirm_required — implausible bearing number, check the plate
Two AIs and a human all read it the same way — and were all wrong. The plate says 6307-ZZ under the glare. A deterministic parts-catalog check flagged it anyway, because 6397 isn’t a bearing that exists. Agreement isn’t truth. That’s why our document reading never silently commits a high-risk number — measured across a 41-image adversarial corpus: zero did.
Fair notes: the transcription above is otherwise correct — the model is genuinely good, which is exactly what makes the one confident miss dangerous. And what our guard can’t catch, we keep on a written ledger instead of pretending. Try this image on your favorite chatbot; ask it for the DE bearing number.
See the box that does this for your paperwork →ThirdShift R&D · thirdshiftrd.com · every claim on this page is from our published test corpus — the numbers live on the receipts page — glare-shot plates, 1936 typewriter invoices, crumpled receipts, and 30 documents built to fool it.