Zero critical defects, fully traced

The problem

A mass warranty incident is never caused by one control failing alone.

A mass, poorly-traced warranty incident destroys trust, triggers field-replacement costs, and puts the OEM customer relationship at risk.

• Every starter battery plant shares the same capital pain: zero critical assembly defects — bad terminal weld, acid leak, faulty lid seal — with full lot-to-battery traceability. Leaks and weak welds that reach the customer are the defects that start warranty campaigns.
• Each control on its own looks acceptable: the marker printed, the camera saw something, the leak tester passed. The defect escapes through the gaps between them. No single control is designed to notice that the others disagree. A leak tester can pass a battery whose date code points to another lot, and nothing today notices.
• A poorly traced warranty incident destroys trust, triggers field replacement costs and puts the OEM relationship at risk. The cost lands on quality, on logistics and on the commercial relationship at the same time.
• And without traceability from lot to battery, the containment covers months of production instead of the lot that actually failed. That is the difference between replacing a few hundred batteries and a few hundred thousand.

How it fits the IRIS system

Agents orchestrating four rings — Connect, Edge and evidence checking each other.

Ring 1 (ERP) defines alloy/lot/model; Ring 2 (Marker) prints with no typing via API; Ring 3 (Edge) reads the assembled unit and cross-checks it against the program — if it doesn't match, JIDOKA AI stops the line; Ring 4 (Evidence pack) ships with every battery. SMED AI speeds up changeovers without breaking the coordination.

No single ring solves the problem; each one exists in some form today and defects still escape. The strength is in the cross-check: when the ERP, the marking, the camera and the evidence stop agreeing, the battery does not ship. That is why the flagship is not a new control, but the coordination of the ones the plant already has.

See the full IRIS architecture →

Before and after

The four battery rings, and what each one closes

Ring	What it does	What it closes
1 · ERP	Defines alloy, lot and battery model for the run	The program every other ring checks against
2 · Marker	Prints serial and date code via API, no typing	Warranty traceability by manufacturing date
3 · Edge	Reads the assembled battery and cross-checks the program	Bad welds, leaks and seals, stopped by JIDOKA AI
4 · Evidence pack	Ships the record with every battery lot	Proof for the OEM and the IATF auditor
SMED AI	Speeds up alloy and model changeovers	Changeovers without breaking the coordination
Human tablet gate	Confirms every captured value	Clean master data behind all four rings

Warranty incidents with mass-claim risk → discrepancy caught on the line, not at the customer.

Impact estimate

Impact estimate — to be validated with your quality lead and plant management.

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 avoided warranty incident.
• Discrepancies caught on the line, not at the customer: the battery that disagrees with its program does not ship.
• Containment scope narrowed from months of production to the exact lot, thanks to lot-to-battery traceability. The heaviest cost is the field replacement campaign, where every battery in the suspect range has to be recovered.
• And protection of the OEM relationship, which is what decides the award of future battery programs.

Estimated payback 6-12 months against the first avoided warranty incident. *Figure to validate* with the quality lead and plant management.

And the fair question from the production manager

“Do we have to deploy all twelve cases to get this?” — no. The rings are built in layers and most cases pay on their own. And when a ring raises a discrepancy, JIDOKA AI stops on a verifiable mismatch between sources, such as a printed date code that differs from the ERP program; reading and comparing those is an anchored task, where the best models drop below 1.5% error [1]. When the doubt is the model's, a person decides.

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