Controlling Penicillium roqueforti inoculation in Roquefort AOP — the mold is the signature, and a signature is not improvised.

The problem

The mold is inoculated in seconds, and shows up two months later.

Roquefort AOP stakes its promise on a narrow margin. Penicillium roqueforti is inoculated in the molding room, before the cheese goes down to the cave. A slightly low dose and the veining does not develop; a slightly high dose and the cheese goes out of balance. The piercing has to be uniform so that air enters every piece equally. And all of that becomes visible not in line, but two months later, when a wheel is cut open in the cave.

In most Roquefort dairies that lives in three separate places:

1. The Penicillium doser — flow, dose and mixing with the curd, often on an old PLC or an analog panel.
2. The piercing machine — pattern, count, depth. Sometimes on old machinery with no interface.
3. The Combalou cave — humidity, temperature, air currents (fleurines). Recorded by the affineur with a notebook and a nose.

The master affineur knows how to read all of it, but cannot be in three places at once. And on the day a curd goes out with a deviated dose, there is no way to know until it can no longer be fixed. The only thing that defends the dairy when a cave deviates is the traceability of what went down there.

How it fits the IRIS system

iLEAN Edge does not touch the cave — it controls what is delivered to the cave.

The problem with inoculation is not a lack of information: it is information that lives on islands and only gets cross-referenced once the damage is already visible. iLEAN acts as the putty that fills those gaps, without asking you to change the doser, the piercing machine or the cave. The Combalou cave does what it has done for centuries; iLEAN makes sure that what goes down to the cave carries the right dose and the right piercing.

Edge sees the dose and the pattern in line. Connect captures the cave environment if it has sensors. The agent cross-references the cahier des charges and holds the curd if the dose deviates. The person signs — never the other way round.

The iLEAN pieces applied to Penicillium roqueforti and Combalou:

• Edge — a terminal with machine vision (CNN) and a flow sensor over the Penicillium roqueforti doser, plus vision over the piercing machine's roller. It reads the dose per curd, the mixing with the mass and the piercing pattern in milliseconds. If anything falls out of range, it triggers an actuator (stack light, doser stop, ejection). It works without a network.
• Connect — captures the natural cave's parameters if sensors are installed, and the affineur's notebooks by photo or by voice. It also captures what arrives from the outside (a change in the internal standard, a notice from the regulatory body) at second zero, without anyone having to forward anything.
• Agent — cross-references the dose, the piercing pattern, the cave environment and the cahier des charges. If a curd has deviated, it does not send an email at 10 pm: it holds the curd and alerts the molding supervisor. The person validates and signs; the curd does not go down to the cave on its own.

See the full IRIS architecture →

Before and after

Manual control vs. in-line control with iLEAN Edge

Aspect	Old doser + notebook + the eye	With iLEAN Edge + Connect + Agent
Penicillium roqueforti dose	An assumption: "the doser is fine"	A reading in milliseconds, per curd
Piercing pattern	Manual sampling of a few pieces	Vision on every piece, pattern verified
Detecting a deviation	Two months later, when the wheel is cut	In line, before it goes down to the cave
Cave environment	The affineur's notebook	Integrated with Connect wherever sensors exist
Operation without network	n/a	Edge keeps running on the panel's own power
AOP file per cave batch	Rebuilt by hand, weeks	A dossier per curd and per cave batch, automatic

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 dairy. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Mid-sized Roquefort AOP dairy, ~1-3 t/day of curd, natural cave in Combalou.
• Edge pilot over the Penicillium doser and the piercing machine (camera + flow sensor + actuator + integration with the MES). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on how often inoculation deviations are detected in the cave and on the average cost of a deviation.
• Expected reduction in scrap from inoculation deviations of ≥ 30%, with a per-curd dossier for the AOP regulatory body.
• The hard lever is a single cave batch not downgraded: given what a Roquefort cave is worth, one of them pays for the pilot.

And the master affineur's reasonable doubt

“What if the camera gets it wrong and lets a badly inoculated curd through?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a sensor reading (reading flow, verifying a pattern), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the doubtful curd and the person signs. The three safety rings exist precisely for this.

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