Zero warranty returns from origin defects

The problem

A safety recall costs far more than the retrieval.

A safety recall isn't just the direct cost of the retrieval: it's the distributor relationship, regulatory exposure and brand reputation in demanding export markets.

• The capital pain of an exporting white-goods plant is a warranty return wave or a safety recall caused by an origin defect: a wrong compressor batch, a mislabeled serial range, an assembly error nobody saw. In a washer or a stove, an origin defect can also become a safety issue in the customer's home.
• The direct cost of retrieving units is only the start. Then come the distributor relationship, regulatory exposure and brand reputation in demanding export markets.
• Each individual control already exists in some form, and each one passes the unit on its own, because each one only checks itself.
• What fails is that no one checks the controls against each other before the appliance is packed and shipped. That cross-check is what the flagship adds on top of the cases it coordinates.

How it fits the IRIS system

Four rings that cross-check — coordinated by Agents, with a human in command.

Ring 1 (ERP: critical-component batch) + Ring 2 (labeler: serial number with no typing) + Ring 3 (Edge: visual reading cross-checked against the order, JIDOKA AI locks the line on mismatch) + Ring 4 (evidence pack: dossier with final test and sign-off). SMED AI keeps changeover speed intact.

No single ring prevents the recall; each already exists in some form today. The protection comes from the cross-check: when the ERP batch, the label, what the camera sees and the test record stop agreeing, the unit does not ship. And a person in quality decides what happens next. The algorithm never releases a unit on its own.

See the full IRIS architecture →

Before and after

The four rings, and what each one secures

Ring	What it checks	What it secures
1 · ERP	Critical-component batch per order	Which compressor or motor went into each unit
2 · Labeler	Serial number with no typing	The serial range matches the order
3 · Edge	Visual reading against the order	JIDOKA AI locks the line on mismatch
4 · Evidence pack	Final test and sign-off	A dossier per unit for any claim
Changeovers in white goods	SMED AI	Speed kept while the rings run
Final decision	Human in command	Quality releases, not the algorithm

Latent untraceable recall risk → in-line discrepancy detection, before the product leaves the plant.

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 6-12 months, measured against the first incident avoided — to be validated with quality and Operations.
• From a latent, untraceable recall risk to in-line discrepancy detection, before the product leaves the plant.
• If a claim does arrive, the affected serial range is bounded by query instead of widened by precaution.
• And protection of the distributor relationship and of the brand in export markets, which no single control secures on its own. Each avoided incident also protects the next export contract.

Estimated payback 6-12 months against the first incident avoided. *Estimate to validate* with quality and Operations.

And the fair question from the production manager

"Will the rings stop the line every hour?" — JIDOKA AI stops on a verifiable mismatch between sources — the label says one model, the camera reads another — not on model doubt. Reading a label or a model plate is an anchored task, where the best models drop below 1.5% error [1]; when the camera is unsure, the unit is escalated to a person instead of locking the line. Stops are rare and justified, and each one is recorded with the evidence that caused it.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.