The Norwegian salmon IQF tunnel — right core, a fillet that sells; wrong core, expensive scrap.

The problem

The core of the fillet cannot be seen — it shows up in the customer's cold store.

The freezing lead in a salmon SBP plant knows the IQF tunnel has three levers pulling at once:

1. Air temperature (inlet and outlet) — the SCADA reads it, but the quality system does not always read the SCADA.
2. Belt speed — adjustable, and under pressure whenever orders peak.
3. Fillet mix — different thicknesses entering at the same time, especially on a multi-SKU line.

The freezing lead knows that if the three do not line up, the core does not reach temperature in time, the ice crystals are large, and the fillet loses water and color on thawing. The problem is that the core is invisible at the outlet: the fillet comes out frozen on the outside and the defect only appears later — in the cold store, at the customer's thawing step, or in the retailer's claim. By the time that claim arrives, a whole pallet has already shipped. The classic system (periodic probe pieces plus a trained eye) works 99% of the time — that 1% is the pallet that comes back.

How it fits the IRIS system

iLEAN does not reprogram your tunnel — it reads the three data points and hands them to the freezing lead together.

The problem is not a lack of information: it is information that lives in islands (the tunnel SCADA knows the speed, the probe pieces know the core of the sample, the order system knows the thickness mix) and nobody cross-references them in real time. iLEAN acts as the putty that fills that gap and gives the freezing lead back the full picture, without asking them to change the tunnel or the SCADA.

Edge measures the probe pieces and models the core of the rest. Connect reads the tunnel SCADA and the production plan. The agent proposes the adjustment. The freezing lead signs — never the other way round.

The three iLEAN pieces applied to the salmon IQF tunnel:

• Edge — a terminal with sensors in periodic probe pieces plus a camera over the tunnel inlet to estimate the thickness and layout of the fillet. A CNN trained to correlate all of it with the core. Calibrated by the freezing lead themselves on your line. It works with no network — if the plant loses WiFi, Edge keeps measuring and storing the data locally.
• Connect — reads the available parameters from the tunnel SCADA (air, speed, alarms), through a modern interface where there is one, OPC-UA if there is one, or a read of the old PLC if that is what you have. And it captures from the production plan the thickness mix expected in each shift. Without asking you to change machines.
• Agent — cross-references estimated core + speed + air + mix, and if the combination falls outside the range that guarantees a good fillet, it proposes the adjustment to the freezing lead through their own channel (traffic light, dashboard, voice). The lead signs the adjustment; the tunnel does not reprogram itself.

See the full IRIS architecture →

Before and after

Probe pieces plus a trained eye vs. cross-referenced core with iLEAN

Aspect	Periodic probe pieces + the freezing lead's eye	With iLEAN Edge + Connect + Agent
Visibility of the core	Only the probe pieces, at isolated moments	Continuous estimate across the whole batch
Change in thickness mix	Noticed late, if at all	The agent cross-references it with speed and proposes an adjustment
Defect discovered	In the cold store, on thawing or in a claim	At the tunnel outlet, before the pallet is packed
SCADA integration	There is none with quality	Connect reads the parameters automatically
Operation with no network	n/a	Edge keeps measuring and modeling locally
File per pallet	The freezing lead's sheet, rebuilt afterwards	Automatic dossier with curves and signatures

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.

• SBP plant with 1-2 IQF tunnels, a mix of SKUs (portions, whole fillet, skinless) and seasonal peaks.
• Edge pilot with sensors in probe pieces + a camera at the tunnel inlet + Connect to the tunnel SCADA + an adjustment agent. First value expected within a few weeks (continuous visibility of the core curve).
• Expected reduction in badly frozen pallets reaching the customer ≥ 30% — a conservative floor.
• Indicative payback between 4 and 9 months, depending on tunnel volume and the average cost of rework/returns. A single pallet recovered before it leaves the tunnel pays for a significant fraction of the pilot.

And the freezing lead's reasonable doubt

“What if the agent proposes slowing the belt down and output drops?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely cross-references structured records (temperatures, speed, thickness) and proposes a value inside the range, the best models brought error below 1.5% ^[1]. And even so, the adjustment is signed by the freezing lead — the proposal arrives as a suggestion, not as an order to the PLC. The three safety rings exist precisely for this.

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