Four rings between the non-conforming part and the road

The problem

Ninety-nine point nine percent ship fine; the whole cost sits in the unit that escapes.

Risk in this sector is not linearly distributed. 99.9% of units ship fine; the entire cost is concentrated in the tail, in the unit that escapes. And what determines the magnitude of the disaster is not only that it escaped, but how long it takes to find out and how precisely it can be bounded. Traceability with gaps turns a suspect lot into a mass recall out of pure caution: you contain everything you cannot rule out. Today four checks exist that, in theory, cover that risk. But each depends on a different person paying attention at a different moment, and none of the four cross-checks automatically against the other three.

• The capital pain of any passive safety plant is always the same: that a non-conforming component escapes. It is not a quality reject, it is a recall campaign across hundreds of thousands of vehicles, and the risk is not linearly distributed — it is all in the tail.
• What determines the magnitude of the disaster is not only that it escaped, but how long it takes to find out and how precisely it can be bounded. Traceability with gaps turns a suspect lot into a mass recall out of pure caution: you contain everything you cannot rule out.
• Four checks exist today that in theory cover the risk: what came in, what the order says, what the line sees, what was recorded per unit. But each depends on a different person paying attention at a different moment.
• None of the four cross-checks automatically against the other three. The escape is never a single failure; it is the concatenation of small gaps that each passed its own control.

How it fits the IRIS system

Four rings that cross-check each other — all of the above, orchestrated.

Four rings, each with its own case behind it:

No single ring solves the problem: each already exists today in some form and the unit still escapes. The strength is in the cross-check, because what gives the concatenation away is four independent sources ceasing to agree — the lot on the container, the variant on the order, what the camera sees on the line and what the unit's record says. If the four do not agree, JIDOKA AI stops the line before the pallet crosses the door, and SMED AI makes sure the rings are not paid for in productivity.

• Ring 1 — Intake (cases 3 and 7): every critical component container enters with its lot and certificate read by photo and validated. If the supplier flags a deviation later, iLEAN knows within seconds which serial numbers carry that lot and where they are.
• Ring 2 — Order and marking (cases 1 and 8): the live order publishes variant, revision and marking format straight to the laser marker, with no keying. The first-off record is captured and validated at start-up.
• Ring 3 — Line (cases 9 and 10): stitching is inspected on 100% of parts by vision before folding; the changeover line clearance is verified camera by camera. On any discrepancy JIDOKA AI blocks the line: it does not warn, it blocks.
• Ring 4 — Per-unit record (cases 5 and 11): every unit ships with its full birth history — lots, torque curve, test result, stitch image, signatures — validated by a human and organized by Agents to be queryable by serial number, by lot or by period.
• SMED AI speeds up model changeovers so the four rings are not paid for in productivity, which is the real objection from the floor.

See the full IRIS architecture →

Before and after

The four rings, and what each one closes

Ring	Cases it draws on	What it closes
1 · Intake	Supplier lot alert (3) + pyrotechnic receiving (7)	Every gas generator lot enters read and validated; a later deviation maps to serial numbers in seconds
2 · Order and marking	First-off record (1) + DMC marker (8)	Variant and revision travel from the order to the marker; the start-up is captured and validated
3 · Line	Cushion stitch (9) + line clearance (10)	100% of seams inspected before folding; changeover verified camera by camera; JIDOKA AI blocks
4 · Per-unit record	Line-side validation (5) + evidence pack (11)	Every unit ships with a human-validated birth history, queryable by serial, lot or period
SMED AI	Runs across rings 2 and 3	Changeovers faster, so the rings cost no productivity
When the rings disagree	All of the above	The line stops before the pallet crosses the door; a person decides

from containment sized out of caution to containment bounded to the affected units · from detection at the customer to detection on the line · from two weeks to answer which units are affected to a query of seconds · from four checks depending on four attentive people to four rings that cross-check each other on their own.

Impact estimate

Impact estimate — to be validated against your containment history and its average cost.

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 against the first incident bounded or avoided.
• The business case is not built on operational saving but on risk avoided: the direct cost of an automotive recall campaign is measured in millions.
• The indirect cost — losing future nominations with that carmaker — is larger still and lasts years, and it is not counted in the payback.
• From two weeks to answer which units are affected to a query of seconds; from four checks depending on four attentive people to four rings that cross-check on their own.

the business case is not built on operational saving but on risk avoided. The direct cost of an automotive recall campaign is measured in millions, and the indirect cost — losing future nominations with that carmaker — is larger still and lasts years. Estimated payback 6-12 months against the first incident bounded or avoided. *Estimate to validate against your containment history and its average cost.*

And the fair question from the production manager

"Do we have to deploy all twelve cases to get this?" — no, and framing it that way would be the way never to start. The rings are built in layers and each case pays on its own from the first month: with pyrotechnic receiving and the first-off record you already have ring 1 and half of ring 2. Every capture underneath is an anchored task, where the best models drop below 1.5% error [1], and every write is validated by a person — the rings are made of the same pieces quality already approved one by one.

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