The Line Won't Start Without a Validated Clean

The problem

The most dangerous moment of the day closes with a signature and nothing behind it.

A line running a seafood croquette and, the next shift, the vegan cauliflower SKU, requires certified cleaning between allergens. Today it rests on a human sign-off with no objective visual evidence.

• A line running a seafood croquette and, on the next shift, the vegan cauliflower SKU needs certified cleaning between allergens. There is no margin here: the vegan and shellfish-free claims are the product.
• Today that transition rests on a human sign-off with no objective visual evidence behind the box that was ticked.
• The residue that matters is small and specific: breadcrumb left in the hopper, batter stuck under the forming table, a trace along the edge of the freezing conveyor belt. None of it is visible from where the sign-off is signed.
• And the changeover happens under time pressure, because it sits squarely on the shift's critical path with the next run and its raw material already waiting at the door.

How it fits the IRIS system

Edge with JIDOKA AI — no start-up without evidence plus a quality sign-off.

Fixed Edge cameras at the critical points — forming table, breading station, freezing conveyor — compare the current state against the 'clean to start' reference. The line won't start until every camera reads OK and the quality lead signs off, now backed by structured visual evidence instead of a blind signature.

The cameras do not decide alone. Each critical point — forming table, breading station, freezing conveyor — is compared against its 'clean to start' reference for the SKU about to run, and the line does not release until every camera reads OK and the quality lead signs. What changes is not who decides, but what they decide on.

See the full IRIS architecture →

Before and after

Today's changeover versus the validated changeover

Aspect	Today	With iLEAN Edge
What closes the cleaning	A signature	Evidence plus a signature
Breadcrumb left in the hopper	Seen if somebody looks	Checked against its reference
Batter under the forming table	Found later, or not at all	Checked before release
Trace on the freezing conveyor	Assumed clean	Checked point by point
Start-up under time pressure	The margin is the sign-off	The line simply does not release
Evidence after a complaint	The signed sheet	Images of every critical point

Cleaning sign-off with no objective visual evidence → Start-up blocked until visual evidence plus quality sign-off.

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.

• Payback scales with the number of allergen changeovers per week, so the estimate is built from your own changeover count rather than offered as a fixed range.
• Direct protection for the vegan and the shellfish-free SKUs, which are precisely the ones whose commercial promise cannot survive a single cross-contamination incident.
• A start-up released on evidence, so the first run of the new SKU stops being born under suspicion and stops being held back as a precaution.
• And an image record of every critical point at every changeover, which is exactly the evidence the audit dossier needs afterwards and the only thing that answers a cross-contamination complaint.

Protects sensitive SKUs · payback scales with changeovers/week. The benefit directly protects the vegan and shellfish-free SKUs against cross-contamination, with payback scaling with the number of allergen changeovers per week. Estimate to validate.

And the fair question from the production manager

"What if a camera blocks the line for no reason?" — that is the right objection, because the changeover sits on the shift's critical path. Each point is compared against the reference state for the SKU about to run and, in case of doubt, the system does not block silently: it escalates to the quality lead, who releases on the evidence. The recognition itself is an anchored task, where the best models drop below 1.5% error [1].

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