Zero off-spec sugar leaving the plant

The problem

The direct cost of the rejection is the least of it. What you lose is the account.

Sugar is an irreplaceable ingredient with no immediate second source for the industrial customer. That is why the direct cost of the rejection is the least of it: what you lose is the account, because a buyer whose line you stopped requalifies the supplier and rarely comes back. There is a structural reason this happens: final quality is decided upstream — incoming raw sugar, affination, carbon dosing, boiling curve, drying — but measured downstream, in the lab, hours later. By the time the result says something drifted out of range, there are tonnes made. Sometimes loaded. And the whole current system depends on one person connecting, in time, signals that live in different systems: an email, a panel, a spreadsheet, a sheet of paper and a coder panel. None of those signals cross-reference on their own.

• Sugar is an irreplaceable ingredient with no immediate second source for the industrial customer.
• That is why the direct cost of the rejection is the least of it: what you lose is the account, because a buyer whose line you stopped requalifies the supplier and rarely comes back.
• There is a structural reason this happens: final quality is decided upstream — incoming raw sugar, affination, carbon dosing, boiling curve, drying.
• And it is measured downstream, in the lab, hours later, with tonnes already made.

How it fits the IRIS system

Four rings: controlled input, process under control, decision, and evidence.

None of the eleven previous cases fixes this on its own. What fixes it is that all four rings share one central memory, so a signal from the certificate at intake can stop a loading at the far end of the plant.

• Ring 1 · Inlet (cases 3 and 7): the raw sugar certificate of analysis and the weighbridge ticket enter on their own and define what raw material is being worked and what color and polarization to expect. If the raw sugar drifts, iLEAN warns before melting.
• Ring 2 · Process (cases 1, 2 and 6): the boiling curve read from the panel, the photographed shift log and the hourly lab analysis cross- reference on the same lot. The deviation is caught while still correctable.
• Ring 3 · Packing and marking (cases 8, 9 and 10): the coder prints what the active order publishes, with no typing; the Edge camera reads what was printed and verifies bag closure; the product changeover only starts with the purge cameras reading OK. If something does not add up, JIDOKA AI stops the line.
• Ring 4 · Release and evidence (cases 5 and 11): nothing is released without a human signature on the tablet summary, and every lot leaves with its evidence pack cross-referencing the previous three rings. If a ring has not closed, loading is not authorized.
• SMED AI accelerates the product changeover so coordinating the four rings is not paid for in productivity — the legitimate objection of any plant manager running an installation designed for maximum flexibility over few machines.

See the full IRIS architecture →

Before and after

The four rings, one by one

Ring	What it secures	With what
1 · Controlled input	Which raw sugar goes into which melt, and what it carries	Receiving by photo + the certificate read as a signal
2 · Process under control	Continuous measurement and dynamic limits	Pan panel, shift log and lab with no lag
3 · Decision and block	Loading authorised, or line stopped	JIDOKA AI + human signature at release
4 · Total traceability	That all of the above can be demonstrated	Agents indexing by certification standard
Coordination	That upstream can stop downstream	One central memory shared by all four
Final decision	Authorise the shipment or stop it	Signed by a person

Impact estimate

Impact estimate — measured against the account, not the lot.

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.

• The calculation is not the cost of the rejected lot, but the value of the account protected.
• An industrial beverage or bakery customer represents a multi-year contract whose value exceeds by orders of magnitude any single rejection.
• Estimated payback 6-12 months, measured against the first off-spec event avoided.
• And the eleven cases before this one stop being eleven separate projects and become one chain.

the calculation is not the cost of the rejected lot, but the value of the account protected: an industrial beverage or bakery customer represents a multi-year contract whose value exceeds any single incident by orders of magnitude. Estimated payback of 6 to 12 months against the first spec incident avoided. *Estimate to be validated* against the rejection history of the last 3 years.

And the fair question from the production manager

"Do we have to implement all twelve cases for this to work?" — no, and trying would be a bad idea. The rings are built in order: controlled input first, because it holds up the other three. Each ring pays on its own; what multiplies is that they share one central memory instead of living in four systems that do not talk.

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