Each contributor photographs one receipt they believe was altered and a genuine one from a similar merchant, boxes the part that looks wrong and writes a line on why. Self-altered receipts, suspects without a comparison and unmarked submissions are excluded, one pair per contributor, so merchants and alteration styles vary across a global base strongest in Southeast Asia, the United States and Europe. The suspicion is the contributor's judgment, not a forensic finding: a weak label with strong localization. Most document-fraud training data is synthetic, and models trained on it risk learning the editing pipeline instead of the fraud; these were found in the wild.