Zero cross-contamination, end to end

The problem

A formula error is almost never a single failure.

A formula, allergen, or batch error in a multi-reference environment (in-house brand plus private label) is the sub-sector's biggest reputational and contractual risk. Each incident costs tens to hundreds of thousands of euros directly, plus potentially irreversible damage to the relationship with the client who buys that formula.

• It is the chain of them: an order misread at start-up, a formula change signed in a hurry, a doser dragging leftovers of the previous filling and a label typed by hand on the packer.
• Each link on its own looks acceptable and passes its individual check. Together they produce the tray that reaches the shelf declaring allergens that are not the ones inside.
• In a multi-reference environment — in-house brand and third-party development on the same lines — that is the sub-sector's biggest reputational and contractual risk.
• Each incident costs tens to hundreds of thousands of euros directly, plus potentially irreversible damage to the relationship with the client who buys that formula. And a chain that pulls a reference over an undeclared allergen rarely gives volume back to the same supplier.

How it fits the IRIS system

Four crossing rings — it is everything above, orchestrated.

The four coordinated rings: the ERP defines the active formula (ring 1); the packer prints exactly what the ERP publishes with no typing (ring 2, legacy integration); the Edge camera reads what's being packed and cross-checks it against the order, locking out via JIDOKA AI if it doesn't match (ring 3); and the evidence pack certifies the three prior rings on every tray (ring 4). SMED AI speeds up formula changeovers so coordination doesn't cost productivity.

No ring on its own solves the problem: each already exists today in some form and the tray still escapes. The strength is in the cross-check, because what gives the chain of failures away is four independent sources ceasing to agree. When the order says one formula, the camera sees another and the packer prints a third, the line stops by itself.

See the full IRIS architecture →

Before and after

The four rings, and what each one closes

Ring	Which cases it comes from	What it closes
1 · Formula	Batch start-up (1) + receiving (7)	Which formula, which allergens and which lots make up the order
2 · Label	MAP packer (8) + tablet validation (5)	Batch, expiry and allergens printed with no typing
3 · Line	Seal defect (9) + allergen changeover (10)	What actually leaves the forming machine, piece by piece
4 · Evidence	Audit dossier (11) + forming machine panel (2)	The proof of the three prior rings, tray by tray
Cross-check between rings	JIDOKA AI	Stops the line when two sources disagree
Formula changeovers	SMED AI	Speeds up the change so coordination costs no productivity

Recurring incidents with partial-recall risk → zero. Discrepancies caught on the line, not at the customer.

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, counted against the first recall avoided.
• Protection against a recall, whose single-event impact can exceed €100,000 between sorting, reverse logistics, destruction and penalties.
• Protection of the relationship with private-label clients, which is what governs contract renewals.
• And the discrepancy is caught on the line, not at the customer: incidents carrying partial-recall risk taken to zero.

Protection against recall (single-event impact potentially >€100K) and against losing private-label client relationships. Estimated payback 6-12 months against the first recall avoided. *Estimate to validate* 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 for itself from the first month. With batch start-up and the packer you already have ring 1 and ring 2, which is where half the allergen risk sits. And in every layer the extractions are anchored tasks, where the best models drop below 1.5% error [1], with human validation before anything is written.

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