What it does well
A fixed overhead camera over a marked tray is good at one plain question: did each expected item show up and settle in the tray? A segmentation model like SAM 3.1 can follow each item, hand and piece of paper frame by frame, and code can count from where things came to rest.
Where it struggles
Be specific about the failures, because they are where a camera needs a person behind it.
- Two items touching can read as one shape. Treat that as a hold.
- Serial labels covered by a finger or paper can be misread. A reader should abstain, not guess.
- Look-alikes fool readers both ways. On ReturnWitness's filmed clips, a reader called a real camera a toy and passed a lens-shaped mug as the lens.
- Anything outside the frame or off the tray isn't seen at all.
What it should never decide
Whether the customer is honest. Whether a part is safe to re-use. Whether to pay. Those stay with a person; the camera's job is to show them the evidence.
How to try one honestly
Start with one bench and the kits that hurt most when a part goes missing. Decide in writing, before you start, what counts as a pass. Report the misses next to the hits, including the complete boxes it held by mistake.
The same routine, with a camera and a person
ReturnWitness counts in app code from where items settle, never from model track IDs, and lets every model reader raise a hold but never clear one. Its pass/fail rule was written down before any run, and that registered test is still incomplete: seven of its clips were never filmed.