Engineered around the constraints, not despite them.

The classification chamber is where the cleverness sits. By fixing the camera, controlling the lighting and taking one bottle at a time, we reduce a hard computer vision problem — object detection in an uncontrolled scene — to an easy one: single-object classification under known conditions. That's what lets a small quantized model run accurately on inexpensive hardware with no accelerator and no cloud call.

Machines that accept a whole bag of mixed material need far more compute, far more mechanism and far more maintenance, to produce a less trustworthy count.

Records are signed on the device the moment a deposit is verified, then written to a local hash-chained store. Each record includes the hash of the one before it, so any later alteration breaks the chain visibly. When connectivity returns, the queue submits to the ledger in order.

Practically: the machine keeps paying, keeps counting and keeps its evidential integrity through an outage. Nothing is lost and nothing can be quietly rewritten in the gap.

We record deposits on Hedera. The reason is narrow and specific: we want the recovery record to be checkable by people who have no reason to trust us.

A database we control is only ever as credible as our reputation. A public consensus record is checkable independently, with timestamps we can't backdate and entries we can't delete. For a claim that will end up in a compliance filing or an assurance process, that difference is the entire point.

We use Hedera specifically for transaction costs low enough to justify writing a record for a single bottle, and finality fast enough to matter operationally.

We are not issuing a token to the public, and there's nothing here to speculate on. The ledger is a filing cabinet nobody can break into.