Three deep categories (power adapters, blender and frappé maker agitators, vacuum
attachments). Other parts go through the generic profile until a profile exists for them.
The app reads what is on the label or in a listing. A wrong or counterfeit label gives wrong
fields; a seller's text is read as written, and only values a page backs count as comparables.
A dimension needs a reference in the same plane; the uncertainty reported is a bound derived
from the marks and the reference, not a measurement of the error. A dimension match filters
candidates and never proves a fit on its own.
Searching and posting on Marketplace stay yours. There is no public listing API for
individuals; the app hands off a search and a create page and never automates facebook.com.
Verdicts are only as good as the evidence. A listing that states nothing ends in
INSUFFICIENT_INFO; a candidate's observation is trusted for a limited time.
The demo's allowances and counts live in one process and start over with it; the spend cap on
disk is the floor.
Without the access code the public demo reads only its own set: a visitor picks the set's
photos and gets the answer recorded for each, and cannot upload one of their own. The code
opens uploads, read by the live models when the host has a key, up to the spend cap the
operator sets.
A blurred or dim label is helped, not fixed: contrast and edges are lifted before the read
and a weak capture is flagged, but the text has to be in the photo; the phone's camera app
focuses where it is tapped.
A clip from the phone is kept and played back; no frame is read from it yet, so what the
label says has to be in a photo.
The decision cards, the uses and a drawing are the model's proposals for your profile, not
checked by the rules: a repair step is advice, and a drawing is an illustration.
An access code, not accounts: every code a deployment accepts opens the same inventory, as
its owner.
Human review is required before anything is posted, bought or sold.