The failure that ruins the year is not scrap

The problem

One human barrier, at the worst moment of the day.

It happens because a dozen brands share the same hall, warehouse and people. The bottles look alike, the caps are interchangeable, the printed material is stacked in the same aisle and the lines change item several times a day. No operator wants to get it wrong. But today the last barrier against mix-up is their attention at the end of a long changeover, with the line stopped and everyone in a hurry. A single, human barrier, at the worst possible moment. And it is a doubly expensive pain: the manufacturer answers financially to its customer and loses the relationship as well.

• A dozen brands share hall, warehouse and people. The bottles look alike, the caps are interchangeable and the printed material is stacked in the same aisle.
• The lines change item several times a day, so the risky moment is not rare: it comes round three times a shift.
• No operator wants to get it wrong. But the last barrier is their attention at the end of a long changeover, with the line stopped and everyone in a hurry.
• And the bill is double: the manufacturer answers financially to its customer and loses the relationship as well.

How it fits the IRIS system

Not a new piece: the eleven previous cases acting at once as four independent rings.

not a new piece, but the eleven previous cases acting at once as four independent verification rings:

The point is not that each ring is perfect. It is that four independent barriers fail together far less often than one tired person at the end of a changeover.

• Intake ring (Connect, case 7): every component and raw material enters with its lot and certificate read off the label, not typed. Whatever is not identified is not available to reserve.
• Changeover ring (Edge + SMED AI, case 10): the line does not start until the cameras confirm nothing from the previous item remains and quality signs. The physical mix-up is blocked before it can exist.
• Start-up ring (Connect, cases 5 and 8): the item-component-lot triad is validated on the tablet with two taps and the coding message flows from the ERP to the coder with no typing, using that brand's template.
• Output ring (Edge + JIDOKA AI, case 9): vision inspects 100% of units and, faced with a systematic anomaly, stops the line rather than keep producing defective product.

Above all four, Agents correlates the vessel curve (case 2), the lab results (case 6), what the supervisor dictated during the shift (case 4) and what the brand flagged by email (case 3); and in parallel assembles the evidence pack (case 11) proving all four rings were active on that specific batch.

See the full IRIS architecture →

Before and after

One barrier vs. four rings

Ring	Today	With the four rings active
Intake	Lots typed in at receiving	Lot and certificate read off the label
Changeover	A signature with no evidence	Start-up locked until every point is green
Start-up	The component confirmed from memory	The triad validated on the tablet
Coding	Typed to each brand's template	Published from the released order
Output	Sampling at final inspection	Vision on one hundred per cent of units
Proving it afterwards	Rebuilt from paper	Evidence generated at the time

from a single human barrier to four independent, verifiable rings · from proving the absence of mix-up by reconstructing paper to proving it with evidence generated at the time · from a recall discovered in the market to an anomaly intercepted on the line.

Impact estimate

Impact estimate — to validate against 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.

• Personal care contract manufacturers where a recall would put the main account under review.
• The direct cost of a single consumer goods recall — reverse logistics, destruction, replacement, contractual penalties — comfortably exceeds deploying all twelve cases.
• Counting only the operational savings of the previous eleven, payback for the set lands between 6 and 12 months.
• The avoided-risk part is the largest and the hardest to put in a spreadsheet: that is a committee conversation, not a cell.

this is the highest-return case and the hardest to quantify up front, because its value is the avoided cost of a rare, catastrophic event. Conservative reference: the direct cost of a single consumer goods recall — reverse logistics, destruction, replacement, contractual penalties — comfortably exceeds the full deployment of all twelve cases. Counting only the operational savings of the previous eleven, payback for the set is estimated at 6-12 months. *Estimate to validate.*

And the fair question from the production manager

“Do all four rings have to be deployed for this to be worth anything?” — no, and trying would be a bad idea. Each ring works and pays for itself on its own; what the four together give is that a mix-up has to get past four independent checks instead of one tired person. The usual order starts with intake and start-up, which are the cheapest and feed everything else, and continues wherever the pain is most expensive.

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