CLP/GHS label verification with machine vision — this SKU's pictograms and H/P statements, not last batch's.

The problem

The wrong label does not look like a defect — which is exactly why nobody stops it.

A specialty chemicals plant juggles hundreds of SKUs, several classification revisions per SKU and export orders in half a dozen languages. The printer pulls its data from the system, yes — but between the system and the container climbing onto the truck there are gaps:

1. Previous batch's labels left in the applicator — after a changeover, labels remain on the roll or in the tray, and the first containers of the new batch go out dressed as the previous SKU.
2. Outdated template — the SKU's classification changed, the label template did not. The whole batch ships with the previous revision's H and P statements, impeccably printed.
3. Destination language — the order goes abroad and ships with the plant's usual language. At destination, that label says nothing to the person who needs to read it.
4. Twin SKUs — two products with nearly identical trade names and different classifications share a line. The operator grabs the roll next door and nobody notices, because the label "looks right".

The classic control is the operator's eye at line speed, and a correct label for the wrong SKU is precisely what the eye does not catch. The defect travels all the way to the customer's receiving dock, the inspection or the incident — the three most expensive places to discover it.

How it fits the IRIS system

iLEAN does not redefine your labels — it verifies that the container leaving the line wears the one your technical team defined.

The classification of each product and the content of its label are defined by your technical or regulatory team under the CLP/GHS framework that applies to your markets. iLEAN does not interpret regulations: it acts as the putty between that definition, the labeling line and the order, and verifies the match unit by unit.

Vision reads every label in line. The agent compares it against the SKU, the revision and the destination of that batch. If anything does not match, the container is set aside. The person signs — never the other way round.

The iLEAN pieces applied to CLP/GHS labeling:

• Vision — a camera over the labeling line. It reads pictograms, H and P statement codes, language, batch number and trade name from every applied label, with a photographic record per container. No need to stop the line.
• Edge — a terminal at the line. It identifies the batch and the order in progress, and shows the operator the expected-versus-read comparison when something is off. It works without a network: if the plant loses its connection, it keeps capturing locally.
• Label agent — cross-references what was read against what is defined for that SKU, revision and destination country. An isolated failure sets the container aside; a repeated failure is treated as a source failure (roll or template) and holds the batch until the supervisor reviews and signs. The person decides — the batch never ships on its own.

See the full IRIS architecture →

Before and after

Eyeball check at line speed vs. label verified with iLEAN Vision

Aspect	Operator's visual check	With iLEAN Vision (unit by unit)
SKU pictograms	Checked "at a glance"	Compared against that SKU's definition, container by container
H and P statements	Illegible at line speed	Codes read and compared against the current revision
Destination-country language	Depends on someone checking the order	Automatic comparison against the order's destination
Previous batch's label after changeover	The first containers slip through	Caught on the first container of the new batch
Evidence of what shipped	None, or an archived sample	Photo of every container's label, tied to the batch
Repeated failure (roll/template)	Discovered after the batch has shipped	Flagged as a source failure, batch held

Impact estimate

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

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Specialty chemicals plant with a broad catalog (hundreds of SKUs), multilingual CLP/GHS labeling and frequent changeovers on the filling lines.
• Pilot on one labeling line (Vision camera + Edge at the line). First value expected within a few weeks.
• Indicative payback between 3 and 9 months, depending on containers/day, number of languages served and the average cost of a label incident (relabeling at destination, returns, holds at inspection).
• Hard levers: label errors caught in line rather than at destination; a per-container photographic record that turns any labeling dispute into a lookup; an end to twin-SKU risk at changeovers.

And the quality manager's reasonable doubt

“What if the vision system confuses a pictogram?” — false positives exist, and that is why the system never decides alone. Hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely compares the pictograms and codes it reads against a list defined for that SKU, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN warns and the operator or the quality manager sign. The three safety rings exist precisely for this.

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