We fed our scanner 102 random photos from Wikimedia Commons

How we ran it

We took 51 files from the Business cards category and 51 from the receipt categories on Wikimedia Commons, downloaded them as they were, and uploaded each one through the public app at app.busytrip.app the way a user would, on 2 September 2026. No prompt tuning between runs, no re-tries. Every scan completed; none got stuck. The counts below come straight from the scan records.

Business cards: 33 read, 18 declined

The 33 that read include cards in French, Spanish, German, Italian, Norwegian, Hungarian and Korean, a Serbian card in Cyrillic and a Russian one printed in pre-1918 spelling. Name, company and title were the fields most consistently filled; email and phone depend on whether the card has them.

The 18 declined split like this:

• 15 were not business cards. A photo of a card case, a closed card box, cards scattered on a pavement, a notice board covered in cards, a sheet of stickers, a leather texture, a CD, a triangular warning sign, two photos of a poster of bank logos, a camera-shop envelope, a wooden card with only a URL, a rap flyer, an old letterhead and an engraved 19th-century trade card with no person on it. The reader answered that it could not find a name, and kept the photo for a retry.
• 3 were real cards it should have read. All three had a company or institution but no person's name: a city council, an antiques shop, and an electronics firm whose card is printed to look like a circuit board. The reader was, at the time, refusing any card without a person. We have since changed that rule: a company-only card is saved with the company as the contact.

Receipts: 36 read, 15 declined

The 36 that read include a Chinese cheese shop, the Dominican tax office, a Hong Kong airport taxi, a Lithuanian supermarket, an Argentine commuter railway, a Polish discount store, a US county treasurer and the US Bureau of Engraving and Printing, plus a run of ordinary till receipts. The amount was filled on all 36; the vendor on 32, with four rows saved as amount-only because no vendor name could be read.

The 15 declined:

• 13 were not receipts. Two pages of a handwritten 19th-century ledger, two inscribed pottery shards, a Japanese IC-chip registration form, a passport entry stamp, a magazine page, a spreadsheet screenshot of an invoice list, a doctor's note, a typed letter, a French proof of vote, a Japanese form and a photo of wax seals.
• 2 were receipts it could not read. An 1880s Austro-Hungarian postal receipt filled in by hand, and a Czech pharmacy slip where the amount is handwritten in the margin. Both are fair misses; a person needs a moment with them too.

Why declining is the feature

A vision model asked to extract a name from a photo of a leather texture can produce a plausible name. That is the failure that hurts: a fake contact in your sheet is worse than an empty row, because you will email it. BusyTrip is built to say Could not read a name from the photo and stop. In this run, none of the 28 non-documents produced a row.

The cost of that design is the five real documents that were declined. Three are fixed by the company-only rule. The two handwritten historical receipts are the honest edge of what a phone scanner should be expected to do.

What we did not measure

This test checks whether the scanner reads or declines. It does not audit whether every field on the 69 rows is exactly right; that needs a labelled set, which we are building. Treat the read rates as a floor for well-formed documents and a ceiling for museum pieces.

Try it on your own stack of cards

Ten scans a month are free, no card required. Open the app, scan the worst card in your wallet, and see what the row looks like.