ZERO ALLERGEN RECALLS

The problem

A recall is almost never one failure: it is four small ones lining up.

The capital pain of any bouillon and soup factory running several brands and formats on the same lines is the same: zero recalls from undeclared allergens, batch or labeling errors. A recall doesn't just cost the recovery operation across hundreds of points of sale — it damages the trust built with the distributor and the consumer, right when the plant is expanding capacity.

• A recipe changed at the last minute, a changeover signed in a hurry, a batch typed into the coder by hand and a pallet released on schedule. Four ordinary Tuesdays in a row.
• Each link passes its own control, which is precisely why the combination survives all the way to the point of sale. Nothing looked wrong at any single step.
• The cost is not only recovering product across hundreds of points of sale: it is the trust of the distributor and of the consumer who read the label and believed it.
• And it lands hardest exactly when the plant is expanding capacity, adding lines and taking on new brands, which is when the coexistence of recipes is at its most complex.

How it fits the IRIS system

Four rings that cross-check — everything above, orchestrated.

The flagship coordinates four rings that cross-check each other: the ERP defines recipe and allergens, the coder prints what the ERP publishes, Edge cameras validate cleaning and batch after every changeover, and the evidence-pack dossier cross-checks the previous three. SMED AI speeds up brand changeovers without sacrificing safety; if anything doesn't match, JIDOKA AI stops the line before the pallet leaves the plant.

No single ring solves it. Each one already exists in some form in your plant and pallets still ship wrong. What gives the chain away is four independent sources that stop agreeing with each other, plus something that stops the line when they do.

See the full IRIS architecture →

Before and after

The four rings and what each one shuts down

Ring	Which cases it draws on	What it closes
1 · Recipe	Recipe startup (1) + receiving (7)	Which formula, which allergens and which raw-material lots
2 · Print	Coder (8) + line tablet (5)	That what is printed is what the order published
3 · Line	Pressing and sealing (9) + changeover validation (10)	The state of the equipment and the piece that actually left it
4 · Evidence	Audit dossier (11) + dryer panel (2)	The proof behind the other three, pallet by pallet
Cross-check	JIDOKA AI across the four	Holds the line when two sources disagree
Brand changeovers	SMED AI between runs	Speeds up the change without spending safety

Yearly incidents with partial-recall risk → zero. Discrepancy detection on the line, not at the point of sale.

Impact estimate

Impact estimate — to be validated with the board, not with production.

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.
• A single recall runs over €100,000 in recovery alone, before you count what the distributor does about it afterwards and what it costs to get back on the shelf.
• Discrepancies are caught on the line, not at the point of sale, which is the whole distance between an internal incident and a public recall.
• And the brand equity of every label you pack, your own or somebody else's, stops depending on whether a particular shift went well.

Protection against recalls (single impact > €100K) and protection of the distributor relationship and the equity of the brands involved. Estimated payback of 6 to 12 months against the first recall avoided. Estimate to be validated.

And the fair question from the production manager

"Do we have to deploy all twelve before any of this works?" — no, and treating it that way is how a plant never starts. The rings are built in layers and each case pays on its own; ring 2 alone, coder plus line tablet, removes the most common cause of a labeling recall. On the extraction every layer rests on, an anchored task, the best models stay under 1.5% error [1], with human confirmation behind it at the line tablet. The plants that get furthest are the ones that ship one ring, measure it, and then decide on the next.

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