Zero Allergen Mix-Ups Across the Range

The problem

An allergen mix-up is never one failure: it is a chain of small ones.

The capital pain of a gourmet croquette plant with 18 SKUs selling into large retail is always the same: zero allergen or labeling mix-ups across the range.

• A sheet from the previous run left on the forming table, a changeover signed in a hurry, a reel of film loaded from the wrong SKU and a label nobody read before boxing.
• Each link looks acceptable on its own and passes its own control. Together they produce the pack that declares the wrong allergens.
• With a range mixing milk, gluten, egg, shellfish and a deliberately vegan SKU, every changeover is an exposure — and short runs multiply the number of changeovers.
• And the cost is not a reject: it is a withdrawal, a retail distributor who stops trusting the brand, and a consumer who bought it directly from you.

How it fits the IRIS system

Connect, Edge and Agents coordinated — JIDOKA AI stops the line on a discrepancy.

The work order and the store order define the SKU and its allergens from the start, already validated by Connect. Batch start-up validates that the recipe matches the planned SKU; the Edge camera over packaging reads what's being boxed and cross-checks it against the active SKU. If it doesn't match, JIDOKA AI stops the line.

No single check solves it: each one already exists today in some form and the pack still gets through. The strength is in the cross-reference, because what gives the chain away is two independent sources ceasing to agree — the label being boxed against the SKU the order planned, the batter made against the recipe the work order called for.

See the full IRIS architecture →

Before and after

The six checks, and what each one closes

Check	Which cases it draws on	What it closes
1 · The order	Batch start-up (1) + online store (8)	Which SKU and which allergens were planned
2 · The recipe	Batch start-up (1) + quality spreadsheet (6)	That the batter made matches the SKU planned
3 · The ingredients	Delivery note (7) + supplier delay (3)	Which lots actually entered the batch
4 · The changeover	Clean validation (10)	That the line was clean for this allergen
5 · The pack	Breading inspection (9) + packaging camera	That what is being boxed is the active SKU
6 · The evidence	IFS dossier (11) + two-tap validation (5)	Who validated each step, and when

Latent risk of allergen mix-up at every SKU changeover → Discrepancy caught on the line, before it ships to retail.

Impact estimate

Impact estimate — to be validated with 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.

• Estimated payback 6-12 months, measured against the first avoided withdrawal and to be validated with management and quality together.
• Protection of the relationship with the large-retail distributor, which is what the IFS Plus certification opened in the first place and what a single incident closes.
• A discrepancy caught on the line rather than on a shelf: the scope goes from a market withdrawal to a run held in the cold store.
• And, in direct sales, a wrong pack that never reaches a consumer's home with your own name on it.

Payback 6-12 months · protection against a product recall. Estimated payback is 6 to 12 months against the first avoided recall, while also protecting the relationship with the large-retail distributor. Estimate to validate with management and quality.

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 is how a project never starts. The checks are built in layers and each case pays on its own from the first month. With batch start-up and the clean validation you already have checks 1, 2 and 4, which is where most of the current exposure sits. Where a model is involved the tasks are anchored, with the best models below 1.5% error [1], and JIDOKA AI stops only on a verifiable discrepancy between sources.

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