Questions stores ask.

Above the returns bench, fixed, pointing straight down. It films the table only: the box, the inspected tray and the hands that unpack.

An item counts once, when it settles in the tray. A hand covering an item, packing paper or an item picked up and put back can't add a second count.

That every part on the packing slip came back. The count is checked against the slip and the store's return policy. Anything missing, two items read as one, an unreadable serial label, or an item that left the tray becomes a hold, never a guess.

SAM 3.1 sees the overhead clip and tracks each item, hand and piece of packing paper frame by frame. Muse Spark 1.3 reads the same evidence a second way (serial labels, damage, look-alike parts). Muse Wonder 1.0 reads the result and drafts the decision against the policy.

In local runs, Meta open-weight models on partner GPUs (Baseten, Modal) audit crops of the tray and read serial labels.

The public site makes no live model calls. It replays a recorded real run, and says so on every result.

On the hosted site: the case record (counts, the reconciliation, the decision trail, who approved and the refund status), which expires after 14 days, and a log of decisions (what was drafted and what the person did), capped at the latest 5,000. Rate limiting keeps a short-lived counter per IP address. Clip upload is switched off on the public site, so it stores no footage.

A pilot request from this site is kept for 90 days so I can reply, and only once storage for it is switched on.

Locally, cases are kept in memory and in a file on the machine running the bench.

For a pilot, what is kept and for how long is agreed with the store in writing before the camera goes up. It isn't decided yet.

A person. They read the evidence and approve with their typed name. That approval is a single-use token bound to that exact return, its evidence and the amount; only then does the refund run.

No model can move money. A model can raise a hold and never clear one, and the decision agent can only move a case toward HELD.

No. A hold says what wasn't observed in the tray, not that anyone lied. The refund waits; it isn't denied. A person reads the clip and decides, and can release the full refund.

Time, not money: the customer's refund waits until a person looks, and that person spends review time on a box that was complete. A backed-up queue means a slower refund, never a wrong one.

It has happened: on one of the four filmed clips, a complete kit was held because a reader called the real camera a toy. It's shown as a loss beside the result.

Four filmed clips so far, which is a count, not a success rate. On the same frames, Muse Spark alone answered 4 of 4 right and the full pipeline 3 of 4, with 0 wrong refunds in both. The one miss was a false hold on a complete kit.

The registered test, whose pass/fail rule was written down before any run, is incomplete: seven of its clips were never filmed.

No. Refunds run in Stripe test mode only, and the server refuses any key that isn't a test key. Merchant, orders and returns on the site are seeded data, labelled as such.

Only when a part is really missing and the refund would have gone out in full. In the example, the $340 kit came back without its $109 lens. A hold that a person later releases saves nothing. The pricing page shows the arithmetic, not measured savings.

Not today. Every refund needs a person.

A release envelope is on the list, not built: rules written in advance that could release a refund under a hard value ceiling, only on complete evidence (the slip matches the box, every item counted, one clean read), with everything else held. The override reasons people type would be watched, and if they turn into boilerplate, the envelope gets tighter.

One kit per store: the app assumes a single kit file, so several kit types at one bench is not built.

The label reader (Muse Glimmer on my GPU) is deployed and measured on held-out labels but not switched on in the app.

Reviewer corrections become training rows for the label reader; that path is built and tested, but no correction has been collected yet.

No accounts, no checkout, no billing. This site states proposed prices; nothing here collects payment details.

The same engine can check a mobility-equipment loan closet's returns: a walker frame, its rubber tips and glides. You can try that setting in the browser.

Its limit, stated every time: it only checks that the parts came back. It never gives medical advice, never judges whether equipment is safe or suitable for a person, and never replaces the closet's own safety check.