Residue / Vault return read

Is this the same card you put in the vault?

One ordinary photo at intake. One ordinary photo on the way out. No lights, no rig, no tag on the card. Every card and every score on this page is real: public scans from the Library of Congress, scored by the actual matcher.

0 false acceptsin 1,222 real pairs of different physical cards, most of them the same card from the same print run

2 of 2 foundcards the Library scanned twice by accident. The matcher found them itself, scoring 68 and 150 against a worst impostor of 11.5

1 second, any laptopno GPU, no cloud, no per-scan fee. A fingerprint is about 2.4 KB

The vault

Twelve cards enrolled at intake. Seven are copies of the same Jack Glasscock card from one 1887 print run. Pick a scenario, or click any card to read it against the vault.

Ready

Pick a scenario on the left.

Score against every enrolled card

Peak-to-sidelobe ratio. Impostors sit near 5. The threshold is 25, twice the worst impostor ever seen.

Two cards the Library scanned twice

We benchmarked on 762 public scans of cards from 1887 to 1914. The collection turned out to contain two single cards the Library had scanned twice and catalogued as two items. We did not know. The matcher found them: identical border wear, identical blemishes, two different files.

Two separate Library of Congress scans of the same Guy Hecker card, side by side

Guy Hecker, catalogue items 0433 and 0434. Two independent scans by the Library's own digitisation, years apart. Score 150. Every other pair of this kind of material scored under 11.5.

Histogram of scores: 1,222 real impostor pairs cluster near 5; the two real duplicate pairs sit at 68 and 150; simulated recaptures spread between

All 1,824 comparisons. Red and grey are real pairs of different cards. Green lines are the two real same-card pairs. Blue outlines are simulated re-photographs, shown as the ceiling they are.

Population	Real on both sides?	Pairs	Accepted at 25
Two different copies of the same card, same print run	Yes	822	0
Two unrelated cards	Yes	400	0
One card the Library catalogued twice	Yes	2	2
Simulated phone re-photograph of the same card	Upper bound	200	156
Simulated bad re-photograph: dark, tilted, blurred	Upper bound	200	55, the rest sent for a retake

How the match works

The thing a person uses to recognise a card is the thing every copy shares. So the matcher removes it on purpose.

1. Align on the printed design, then remove it

The printed pattern is the strongest signal in the frame, which makes it a perfect alignment target and a worthless identity signal. It lines the two photos up, then it is subtracted.

2. Compare the surface texture

Paper fibre, where each ink dot landed, a century of wear. What is left after subtraction is correlated. A real match is a sharp peak; a different copy is flat noise.

3. Match, no match, or retake

Match, no match, or retake the photo. A tool that cannot tell "different card" from "bad photo" turns every blurry picture into evidence of fraud. This one says so and asks for another shot.

4. Runs on your own machine

Two Python dependencies, no trained model, no network call. A registry is one file. The whole benchmark reruns with one command.