The line won't start without the cameras' green light

The problem

The checklist is well designed. The problem is how it gets signed.

Running several formulas and allergens on the same lines — in-house brand plus contract manufacturing — multiplies critical changeovers per shift compared to a single-formula plant. Today it depends on someone signing off 'clean' with no objective visual evidence, and leftover filling from the previous run can contaminate the first batch of the new formula.

• Going from a formula with dairy to one without requires certified cleaning of hopper, doser, conveyor and cutting tooling. Leftover filling from the previous run contaminates the first batch of the new formula.
• Running several formulas and allergens on the same lines multiplies critical changeovers per shift compared to a single-formula plant.
• Today everything depends on somebody signing “clean” with no objective visual evidence behind the box, and it gets signed with the line stopped and the shift clock running, which is the worst possible combination for a careful check.
• When an incident appears, the investigation runs into a ticked box and no proof. It is not bad faith: a signature is not a check, and the private-label client who audits knows it perfectly well when they ask to see the evidence for the last thirty changeovers.

How it fits the IRIS system

Edge with JIDOKA AI — no start-up without conformity and sign-off.

Edge plus JIDOKA AI: fixed cameras with CNNs trained to recognize "clean" versus "leftover filling or dough from the previous run" at each critical point. The line won't start until every camera clears and quality signs off based on visual evidence.

The camera does not decide alone: it compares each critical point against its 'clean' reference state, and it is the quality lead who signs on that evidence. What changes is not who decides, but what they decide on. And along the way you are left, at no effort, with the photo the auditor will ask for six months later.

See the full IRIS architecture →

Before and after

Today's formula change versus the validated change

Aspect	Today	With iLEAN Edge
Cleaning sign-off	A ticked box	Critical points checked against reference
Leftover filling from the previous run	Found in the first batch	Blocks start-up
Cutting tooling and doser	Eyeballed	State compared image to image
First batch of the new formula	Under suspicion	Validated before the first disc
With many changeovers per shift	More accumulated exposure	The same check on every one
Evidence for the private-label client	Nonexistent	Photo and sign-off per changeover

Blind sign-off with cross-contamination risk → sign-off based on visual evidence at every critical point.

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.

• Protection of private-label contracts: the brief sets no payback range here, because the cost of losing a client over an allergen incident far exceeds the cost of the cleaning itself.
• The blind sign-off becomes a sign-off based on visual evidence, critical point by critical point.
• The risk of cross-contamination between formulas with and without dairy stops depending on the shift's haste.
• And for every formula change the plant ends up with visual proof to show the private-label client who audits before renewing the contract.

Protects private-label contracts — losing a client over an allergen incident costs far more than the cleaning itself. *Estimate to validate*.

And the fair question from the production manager

“What if the camera blocks the line for no reason with the shift clock running?” — that is the right objection, because the formula change sits on the critical path. Conformity is checked against that specific point's reference state and, in case of doubt, the system does not block silently: it escalates to the quality lead, who signs on the evidence. It is an anchored task — fixed scene, two known states — where the best models drop below 1.5% error [1]. It stops on a checkable discrepancy, not on model uncertainty.

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