Find out here, not on the customer's line

The problem

Approval takes years to build and is lost in a week.

When a batch with an out-of-range melt flow index, with cross-contaminated resin or with black specks reaches an injection moulder or a thermoformer, the customer's line stops. The whole truck comes back. A non-conformity is opened. And if the customer is pharmaceutical or demanding manufacturing, the approved-supplier status itself is at risk, which takes years to build and is lost in a week.

• When a batch with an out-of-range melt flow index, cross-contaminated resin or black specks reaches an injection or thermoforming machine, the customer's line stops.
• The whole truck comes back. A non-conformity is opened.
• And if the customer is pharmaceutical or a demanding manufacturer, supplier approval is put at risk.
• Approval that takes years to build and is lost in a week.

How it fits the IRIS system

Four verification rings running from the bale to the big bag.

The flagship system coordinates the eleven previous pieces into four verification rings that run from the bale to the big-bag. The first is inbound: the real blend is recorded from the weighbridge and the blend sheet, so we know what every kilo is made of. The second is process: the pelletiser curve is read off the panel and tied to the batch, so deviation is visible while it is happening. The third is product. The Edge camera controls black specks, gels and colour continuously over the belt, and purge validation protects start-up after every changeover. If the defect crosses threshold, JIDOKA AI diverts the material and halts bagging. It does not alert somebody so they can decide: it stops the line, and then it alerts. The fourth is release, and it is the lock. The lab result reaches the ERP on its own, and the big-bag label is issued from the ERP. A batch with no validated result cannot be labelled, and what is not labelled does not ship. The auditable pack travels with the shipment. And SMED AI orders the grade changeovers so that the demands of the four rings are not paid for in productivity. Above the four rings sits a fifth thing, and for the P and L it may be the main one: live mass balance. Kilos in across the weighbridge against kilos of saleable pellet out, per batch and per shift, on the screen of the person who can correct it during the shift rather than the following month. In this business yield is margin, and today it is known once nothing can be done about it.

None of the eleven previous pieces solves this alone. What solves it is that all four share one central memory: a signal at intake can hold a big bag at dispatch.

See the full IRIS architecture →

Before and after

The four rings, one by one

Ring	What it secures	With what
1 · Intake	What every kilo is made of	Weighbridge and blend sheet, with the real blend
2 · Process	Seeing the deviation while it happens	The pelletiser curve tied to the batch
3 · Product	Black specks, gels and colour	Continuous Edge over the belt
4 · Evidence	That all of the above can be shown	Agents indexing by standard and period
Coordination	That intake can stop dispatch	One central memory shared by all four
Final decision	Release the big bag or hold it	Signed by a person

Impact estimate

Impact estimate — measured against the rejection, not the batch.

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 batch rejection avoided.
• And yield loss becomes visible by batch and by shift, where today it is a monthly figure with no assignable cause.
• Less exposure of your supplier approval, which is what is really at stake with a demanding customer.
• And the eleven pieces before this one stop being eleven separate projects and become one chain.

Payback 6-12 months against the first rejection avoided · and yield becomes visible per batch and shift. Estimated payback is six to twelve months measured against the first batch rejection avoided, and that only counts freight, material, rework and the customer's downtime. Protecting approved-supplier status is another order of magnitude and it is multi-year. An estimate to validate against the rejection history and the plant's real yield. Conservative ranges, an estimate to validate against the plant's real data.

And the fair question from the production manager

"Do we have to implement all twelve for this to work?" — no, and trying would be a bad idea. The rings are built in order: intake first, because it holds up the other three. Each ring pays on its own; what multiplies is that they share the same central memory.

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