The line does not start until Edge and quality give OK — the critical changeover in a processed cheese plant.

The problem

The critical changeover is where the seal is at stake, not steady-state production.

In a dairy processing plant sharing machines between a lactose-free line and a lactose line — or between PDO and non-PDO product — the running process is stable. The risk concentrates in the critical changeover: the moment the same slicer, the same packing machine or the same thermoformer switches from one product to another and everything depends on the cleaning and the material change having been perfect.

• The lactose-free certification — worth gold commercially, and today it hangs on a human signature with no objective visual evidence: the quality manager signs that the cleaning is OK because the protocol was followed, not because they could check every critical point with their own eyes.
• The PDO certification — requires that the package, the label and the format correspond exactly to the designation at every SKU change. A mismatch is not just a labeling error; it is a non-compliance before the PDO certifier.
• A single minimal residue between cleanings — a trace on a cutting guide, a remnant at the thermoformer's mouth — contaminates the new product's first batch before anyone notices.

Cross-contamination or a PDO mismatch translate into the same thing: a recall, a sanction from the relevant certifier and, in the case of the lactose-free seal or the protected designation, the end of that certification for the plant. A seal built over years can be lost in one badly verified shift change.

How it fits the IRIS system

Edge gives the OK point by point — JIDOKA AI does not let the line start until everyone has said yes.

The critical changeover is not solved with a longer checklist for the quality manager: it is solved by giving them objective visual evidence of each point before they sign the release, and by making the line physically unable to start while that evidence is incomplete. That combination is Edge plus JIDOKA AI.

Edge looks at every critical point of the changeover — cleaning, guides, packing, thermoforming. JIDOKA AI does not release the startup until every camera gives OK. Quality signs on the visual evidence, not blind.

The iLEAN pieces applied to the critical changeover in a processed cheese plant:

• Edge — fixed cameras at the changeover's critical points: the slicer's infeed belt, the cutting guides, the packing room, the thermoformer's mouth and a final camera on packed product. Each camera learns that point's reference state — clean for the lactose changeover, set to the correct format for the PDO/non-PDO changeover — and compares against it at every batch change.
• JIDOKA AI — receives each camera's verdict and acts as the startup gate: while a single critical point is not OK, the line stays held. It is not an alert someone can ignore under shift pressure; it is an active block of the startup until the evidence is complete.
• Quality's signature on evidence — when every camera gives OK, the quality manager receives the summary with the photos of each critical point and signs the release. The signature remains human and remains mandatory — what changes is that it is no longer signed blind, it is signed seeing the evidence backing each point.
• The changeover dossier archived — the photo of each critical point, Edge's verdict and quality's signature are archived per batch change, ready for the IFS/BRC auditor or the PDO certifier with nobody having to reconstruct anything weeks later.

See the full IRIS architecture →

Before and after

A blind signature vs. a signature on visual evidence per critical point

Aspect	Blind signature	Signature on visual evidence
Verifying the cleaning between lactose and lactose-free	By eye, per protocol, no objective evidence	AI vision at every critical point, with an archived photo
Verifying the PDO / non-PDO changeover	Trust that the operator loaded the right material	A camera validates label, mold and format before starting
Quality's release signature	Blind — signed with no visual evidence of each point	On point-by-point visual evidence, in front of the manager
The line's startup	Depends on judgment and the shift's pace	JIDOKA AI holds it until every camera gives OK
A failed camera or missing reading	Assumed OK if nobody says otherwise	Block maintained — the absence of data is not a green light
The file for the auditor or certifier	Rebuilt by hand after the incident	A per-changeover dossier: photo + verdict + signature, archived

Impact estimate

Impact estimate for your plant — to be validated with your own 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.

• Dairy processing plant with machines shared between a lactose-free line and a lactose line, or between PDO and non-PDO product, with several critical changeovers per week.
• Edge + JIDOKA AI pilot on the priority changeover's critical points (cameras + startup gate + integration with quality's signature). First value expected within a few weeks.
• What is protected is the revenue tied to the lactose-free seal and the PDO designation — in many plants counted in hundreds of thousands of euros per product segment. A single seal loss or a single recall compromises that entire volume, not only the affected batch.
• The payback is a function of the lactose-free/PDO mix's weight in your plant: the higher the share of certified production on shared machines, the sooner the pilot pays off. Estimate to be validated with your data.
• The hard lever is a single certification loss avoided: the recall, the certifier's sanction, the rework and the time — months or years — it takes to recover a lost seal.

And the fair question from the quality manager

"What if the camera gives OK to a point that actually is not, or the reverse, blocks the line for no real reason?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI limits itself to comparing a critical point's current image against its previously validated reference state — which is exactly what Edge does here — the best models brought the error below 1.5% ^[1]. And even then, nothing critical is decided alone: JIDOKA AI holds the startup and the person signs, seeing the evidence backing each point. iLEAN's three safety rings are there precisely for this — the AI proposes, the quality manager decides.

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