Skip to content
ReturnWitness
Returns bench · kit completeness

AI can now see what's in a returned box. Should it be allowed to pay you?

  1. A box comes back. A scan says "received". The refund goes out, even if the lens is still sitting at home.
  2. That's $340 refunded for $231 of kit, and the next customer opens a "complete" kit with no lens.

ReturnWitness changes that.

Not AI deciding for us. AI showing us the evidence.

Return a box yourselfWatch a refund get heldRequest a pilot

The store, the customer and the return are a seeded scenario. Try-it runs in your browser: app code, no model.

RMA-20418 · Creator Kit · $340

Diagram of RMA-20418: the box held camera body and battery + charger. 2 of 3 kit items were counted in the inspected tray; the zoom lens line reads 0 of 1. The whole $340 refund is held and waits for a person to decide.

Diagram of the seeded return, not model output. The Instant Demo replays the recorded run and labels every result.[Seeded merchant order & RMA fixture]
The problem
In this scenario, a “received” scan would release the refund, not what was in the box, so an incomplete kit refunds in full.
The solution
SAM 3.1 tracks each item, app code counts what settles in the tray against the packing slip, and the refund holds until a person approves.
Why it matters
The missing $109 zoom lens is caught at the bench, with a timestamped record, not at a restock count later.
How it’s used
Pick the RMA, unbox one item at a time into the tray, read the lines, then approve or keep the hold.
The problem

Rita's afternoon at Bench 2.

Rita Calder runs returns at Brindlewood Camera Co., an online camera store. The Creator Kit is a camera body, a zoom lens and a battery with charger: $340 together.

Illustrative workload, not measured

  • 60–80 returned boxes on her bench each afternoon
  • about 40 seconds per box for the person opening it

A picture of the job for this scenario, not a count from a real store.

Picture a full bench of returned boxes every afternoon and little time for each. The customer on RMA-20418 says “everything's in there.” She lifts out the camera and the battery, a wad of kraft paper, and the box looks empty enough.

In this scenario, the return system marks the RMA “received” on the scan, and that scan would release the $340 refund on its own that night. The lens was never in the box. Someone finds out at a restock count a week later, and by then nobody can show whether it went missing at the bench or on the shelf.

Why it matters. The store absorbs the $109 lens, or the whole kit if the rest can't be resold, and it has no record to dispute with. The person at the bench has no time and no evidence. The gap is that the refund follows a scan, not the contents.

Who it's for

Returns leads at online stores that sell multi-part kits.

If a box that comes back short still refunds in full, the refund is following the scan, not the contents. That's the job ReturnWitness takes on.

Camera and lens kits

A body, a lens, a battery and charger: small, expensive parts that are easy to leave at home.

Drones

Spare batteries, propellers, the controller: a kit is only complete with every part.

Power-tool sets

The tool, its battery, the charger, a guard: a set sold as complete should come back complete.

Toys with parts

Building sets and play kits, where one missing piece means the box can’t go back on the shelf as complete.

Before · the scan decides

The box is scanned. The refund goes out.

  1. 01The return is scanned: “received”.
  2. 02The $340 refund releases on the scan. refunded in full
  3. 03The $109 zoom lens was never in the box. The short kit goes back on the shelf, with no record of what was seen.
After · the count decides what waits

The camera counts. A person decides.

  1. 01The box is unpacked into the tray under the overhead camera.Camera body1 of 1Zoom lens0 of 1 · not observedBattery + charger1 of 1
  2. 02$340HELDThe policy clause is quoted beside it.
  3. 03A person reads the evidence and decides: hold, a partial refund under the policy, or the full refund.

A second setting, same engine: a mobility-equipment loan closet checking that a walker came back with its rubber tips and glides. It only checks that the parts came back; it never judges whether equipment is safe for a person.

Try the loan-closet benchWho it's for, in detailReturn a box yourself
The solution · how it's used

Count, reconcile, hold, approve, refund.

The camera watches. Code counts. Models read and draft. A person decides. The product is the workflow around the refund, not a picture of a box.

  1. Step 1: Film and segment

    The overhead camera films the unboxing. SAM 3.1 segments each item and keeps following it from frame to frame.

  2. Step 2: Count in code

    An item counts once, when it settles in the inspected tray. That rule lives in app code, built not to add a count for hands, kraft paper or re-issued track IDs; it passes the synthetic stress sweep, and in the reported run it counted every item right; on g1 it also flagged one mask covering more than one item (a merged-mask hold), and the refund was still recommended for a person to release; the registered gate is incomplete, g2–g8 not filmed.

  3. Step 3: Reconcile and hold

    App code reconciles the counts against the packing slip and the policy. When enabled, Muse Spark 1.3 reads the same evidence and can only add a hold, never clear one.

  4. Step 4: Draft, approve, refund

    Muse Wonder 1.0 drafts the decision and cites the policy clause. A person approves. Then a Stripe test-mode refund runs.

Read the full pipelineSet up a bench
The held return

RMA-20418: the lens was not observed.

The card is the packing-slip rule as a diagram: two lines seen, one missing, so the whole $340 is held and the policy clause that applies is quoted next to it.

The Instant Demo replays the recorded filmed run of this return. There the lens was not observed, and the camera line was held too, on a false stand-in flag:

Loss misread: the prop is a real compact film camera, not a toy (it stands in for the kit's interchangeable-lens body); a false stand-in flag

The Instant Demo plays a recorded result. It does not make a live Meta call, and the label on the card says so. The merchant, the customer and the RMA are seeded data.

The packing-slip rule, as a diagram (not the recorded run)

RMA-20418 · Creator Kit

Brindlewood Camera Co. · Bench 2 · order ORD-58213
HELD
Expected versus seen settling in the tray, per packing-slip line
Packing-slip lineSeen / expectedResult
Camera body
CK-BODY-01 · $199
1/1OK
Zoom lens
CK-LENS-24105 · $109
0/1MISSING
Battery + charger
CK-BATT-CHG · $32
1/1OK
Refund requested ($340)$340 held
CLAUSE 4.2 · COMPLETE KITS

“A kit is refunded in full only when every component listed on the packing slip is returned.”

RECORDED FIXTURE · synthetic scene, not a live Meta call[Seeded merchant order & RMA fixture]

What a HELD shows, and what it doesn't

What the record shows

No accepted arrival of that item in the tray during this clip, or something the camera couldn't resolve: two items merged into one, a label it couldn't read. A hold is a question for a person, not proof the customer kept anything.

What it doesn't show

That the customer kept the item, that an item it did see is genuine, or that there is no hidden damage. A brick the shape of a lens is out of scope, and so is anything inside a sealed box.

So it never denies

It holds the refund, attaches timestamped evidence and the policy clause, and a person decides. Under clause 5.1 a hold is not a denial.

Said about it · not a customer: no store uses ReturnWitness yet

“What you're really selling is an auditable hold, not a cautious AI.”

“Then watch whether her override reasons are ever boilerplate. If they are, tighten the envelope, not the person.”

Boardy, @boardyaiIn a direct message about ReturnWitness, quoted with permission.
Results · 4 filmed clips · losses firstRECORDED · one live run

Same frames, two answers (n = 4)

Muse Spark alone 4/4 · the pipeline 3/4 · 0 wrong refunds in both

d2Complete kitHELD: a false hold · a loss
g0Empty trayHeld, correctly
g1Clean unpack (the calibration clip)Refund recommended, correctly (a person releases it)
d1Zoom lens missingHeld, correctly

These are counts on 4 clips, not a success rate. The pipeline arm wasn't run the same way on every clip, so its score is flattering. The registered gate is INCOMPLETE: g2, g3, g4, g5, g6, g7, g8 not filmed.

Every number on the evals page
Proposed pricing · no store pays this yet

Start with one bench for 30 days.

Pilot · one bench, 30 days: free, up to 1,000 inspected returns. Then $2.00 per high-risk return inspected.

Break-even illustration, before other costs: One really-missing $109 zoom lens not refunded covers 54 inspections at $2.00.A hold that a person later releases saves nothing. This is arithmetic, not measured savings.

See the options and the arithmeticRequest a pilot
Proposed · per high-risk return inspected
$2.00
Questions stores ask

Where the camera sits, what it keeps, who decides.

All questions
Where does the camera sit?

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.

Who decides the refund?

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.

What does a false hold cost?

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.

Can it release a refund on its own?

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.

Guides for returns teams

Practical guides.

How to check returned items for missing parts

A bench routine for multi-part kits: a checklist per SKU, one place to count, and a hold instead of a guess.

Holding a refund without accusing the customer

How to word a hold so it states what was observed, keeps the refund waiting rather than denied, and puts a person in charge.

What a camera can and can't see at the returns bench

What an overhead camera and a segmentation model are good at on a returns bench, where they fail, and how to design around it.

Put one bench on it.

Tell me what kits you sell and how many come back a week. Or return a box yourself first: it runs in your browser, test mode only, no real funds move.

Request a pilotReturn a box yourself