Get a structured, honestly-flagged read on an uploaded image

Why this works

Vision models are documented to hallucinate plausible values for illegible or occluded regions of an image rather than reporting uncertainty by default, because the underlying generation objective favors a fluent, complete-looking answer over an honest gap, the same fluency-over-accuracy tendency text generation has, applied to pixels instead of facts. Instructing it to flag illegibility per item, rather than trusting a single end-of-response caveat, keeps the uncertainty attached to the specific value it belongs to instead of a blanket disclaimer that does not tell you which number to actually go double-check. The rule against inferring a typical value targets a specific vision failure mode on documents like receipts and invoices: the model has seen enough receipts with a printed subtotal that summing line items to a plausible total, when no total field is actually shown, is a completed pattern it is inclined to finish even though nothing asked it to compute anything. Asking for all matches rather than the single most likely one prevents silent disambiguation on an image with, say, two dates or two totals visible, where a confident wrong guess is worse than a surfaced ambiguity you can resolve yourself in two seconds.