Card identification runs on our server. When you scan, the app crops the photo to the card, reduces its size and uploads it. We read the card’s text, check whether it is a graded slab, and match it against our catalog. Scanning needs a connection. Photos you choose from your library are handled the same way, and the app reads only the photos you pick.
We keep scan photos to improve identification. Each cropped photo is stored with what our system read from it and the result, plus a one-way hash of your account id instead of the id itself. If you correct a result, the correction is stored against that scan. These photos are used to measure and train the models that recognise cards. They show the card you scanned, so frame the card and keep other things, such as faces or documents, out of the shot.
When our own matching is not confident, the cropped photo and up to five candidate cards may be sent to a third-party AI vision service to choose between them. The provider processes that request only to return an answer, and under its commercial terms does not use it to train its models.