Four rings that verify each other, from recipe to pallet

The problem

An allergen or mislabeling incident is never a single failure.

What makes this pain critical is the combination of four conditions: a broad catalog, shared lines, a dense allergen matrix (milk, gluten, egg, fish, crustaceans, molluscs) and multi-language labeling for export. Each one on its own is manageable without drama. All four together turn every working day into dozens of critical transitions, and only one has to fail: a run contaminated with crustacean, an export case carrying the allergen declaration of the previous recipe. And the cost that never appears on the recall invoice is the highest one: the relationship with the distributor, worth considerably more than the incident.

• What makes this pain critical is the combination of four conditions: a broad catalog, shared lines, a dense allergen matrix — milk, gluten, egg, fish, crustaceans, molluscs — and multi-language labeling for export.
• Each one on its own is manageable without drama. All four together turn every working day into dozens of critical transitions, and only one has to fail: a run contaminated with crustacean, an export case carrying the allergen declaration of the previous recipe.
• One incident costs the recall, the distributor penalty, collection across hundreds of points of sale and, in the worst case, the contract.
• And the cost that never appears on the recall invoice is the highest one: the relationship with the distributor, worth considerably more than the incident.

How it fits the IRIS system

Agents coordinating four rings that cross-check — it is all of the above, orchestrated.

Four rings, each with its own case behind it:

No single ring solves the problem: each one already exists today in some form and the incident still gets through. The strength is in the cross-check, because what gives a bad transition away is four independent sources ceasing to agree — and when they disagree, JIDOKA AI stops the line before palletizing.

• Ring 1 · The recipe (Connect, cases 1 and 5): the active order and its recipe sheet define what is produced, from which raw materials and which allergens it declares. Captured by photo and validated by a human on the tablet.
• Ring 2 · The changeover (Edge + JIDOKA AI, case 10): the line does not start the new reference until the cameras confirm the critical points are clean and quality signs on that evidence. SMED AI accelerates the changeover sequence so that rigour does not penalize productivity.
• Ring 3 · The mark on the product (cases 8 and 9): the coder prints lot, expiry and allergens straight from the order, with no typing; and the line camera reads back what is printed and cross-checks it against the active order. If they do not match, JIDOKA AI stops the line before palletizing. Two independent systems checking each other.
• Ring 4 · The evidence (Agents, case 11): every batch leaves with its dossier — recipe, cleaning evidence, tunnel curve, laboratory controls, photograph of the code and signatures — organized by standard and available with no human work.

See the full IRIS architecture →

Before and after

The four rings, and what each one closes

Ring	Which cases it draws on	What it closes
1 · The recipe	Batch start (1) + line-side sign-off (5)	What is produced, from which lots, declaring which allergens
2 · The changeover	Allergen changeover (10)	The line does not start until every critical point clears on evidence
3 · The mark on the product	Coder (8) + coating vision (9)	Printed lot, expiry and allergens read back against the active order
4 · The evidence	Evidence pack (11)	Every batch leaves with its dossier, organized by standard
When two rings disagree	JIDOKA AI	The line stops before palletizing
When the customer calls	Agents	The batch dossier, in minutes

Impact estimate

Impact estimate — to be validated with quality and 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 incident avoided.
• The direct cost of that incident runs into tens or hundreds of thousands across recall, penalty and collection from hundreds of points of sale, without counting the distributor relationship.
• The discrepancy is caught on the line instead of at the customer, and an alert is answered with the batch dossier in minutes instead of after days of reconstruction.
• And SMED AI accelerates the changeover sequence, so the rigor of ring 2 does not penalize productivity.

the return is measured against the first incident avoided, whose direct cost runs into tens or hundreds of thousands, without counting the distributor relationship. Estimated payback of 6 to 12 months. *Estimate to be validated with quality and management*.

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. With the batch start, the line-side tablet and the coder you already have ring 1 and half of ring 3, which is where mislabeling is decided. Each ring rests on anchored tasks where the best models drop below 1.5% error [1]; what the cross-check adds is that a failure in one ring is caught by another before the pallet leaves.

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