The evidence pack of the cleaning between gluten / gluten-free batches — the AOECS audit stops being a sprint.

The problem

The annual sprint before the auditor — and the batch nobody can reconstruct when an alert arrives.

In a breadstick or crispbread factory sharing a line between wheat product and gluten-free product, the operational reality of the cleaning between batches usually looks like this:

• The operator takes photos with their personal phone at every cleaning — they sit in their camera roll, mixed with their birthday pictures. When they go on vacation, those photos are nowhere in the plant.
• The shift lead fills in an Excel sheet at the end of the shift, ticking "cleaning OK" in a box. The signature is after the fact, sometimes the next day, almost always without going back to the station to verify.
• Three weeks before the AOECS, IFS or BRC audit the sprint starts: the quality manager traces photos on phones, reconstructs Excel sheets, chases pending signatures and assembles a dossier that looks coherent. Two weeks of extra work for a person who was already stretched.
• And if mid-year a health alert arrives — a celiac consumer reporting a reaction, an analysis testing positive on a specific batch — the authority's question is very specific: "show me how the line was cleaned on the day of batch 247-B". Reconstructing that specific change, weeks later, is often impossible. And not being able to prove it equals having done it wrong.

The classic system works 99% of the time. That 1% is what puts the Crossed Grain mark on the package at risk — and with it, the contract with the retailer that requires it.

How it fits the IRIS system

iLEAN does not add a fourth system — it turns the cleaning into a file that closes itself.

The evidence pack's problem is not that the cleaning is done badly: the problem is that the proof it was done right lives in places incompatible with an audit (personal phones, Excel sheets, the shift lead's memory). iLEAN acts as the putty that collects that proof at the exact moment it is being generated, packages it, signs it and deposits it in your QMS. Without anyone changing how they clean.

Edge captures the pre/post photo at each critical point. Agents assembles the per-batch-change dossier with a timestamp, the responsible person and a hash. The person signs on the spot — the file is closed before the next batch. The audit stops being a sprint.

The two iLEAN pieces applied to the cleaning evidence pack:

• Edge — a machine vision terminal (CNN) that runs a pre-programmed walk through the breadstick/crispbread line's critical points: the mixer, the flour hopper, the molding belt, the oven, the cooling belt, the packing exit. At each point, a pre-cleaning photo and a post-cleaning photo with a timestamp. It works without a network: if the plant loses WiFi, Edge keeps capturing and storing locally; when the network returns, it syncs. The critical part cannot depend on connectivity.
• Agents — an agent collects Edge's images, crosses them with the current cleaning SOP, the order change recorded in the MES (which batch ended, which starts) and the signature of the person who validated on the spot, and packages the per-change dossier as a PDF with an integrity hash. It deposits it in the QMS or the repository you already use. The file lives in your system — iLEAN does not build a fourth silo, it feeds the one you have.

See the full IRIS architecture →

Before and after

The annual sprint vs. a file closed before the next batch

Aspect	Cleaning documented by hand	With iLEAN Edge + Agents
Capturing the photo	The operator's personal phone, a private camera roll	An Edge terminal on the floor, a pre-programmed walk, a central archive
The AOECS / IFS / BRC pre-audit	A 2-3 week sprint reconstructing dossiers	Zero. The file is already closed and signed per change
The responsible person's signature	After the fact, sometimes the next day	At the moment, on the spot, anchored to the photo pack
Reconstructing a specific batch months later	Hunting photos on phones, recalling who was in that day	Filter by batch number → a signed PDF in 3 clicks
Operation without a network	n/a (the operator's phone does not depend either, but their camera roll does)	Edge keeps capturing on the cabinet's power; syncs on return
The file's integrity	An Excel sheet editable after the signature	A PDF with a cryptographic hash — tampering shows

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.

• SME-type breadstick/crispbread factory, 1 shared line with daily or weekly gluten / gluten-free batch changes, a current AOECS certification plus IFS Food.
• Edge + Agents pilot on that line (a camera terminal at the 5-6 critical points + integration with the MES/QMS). First value expected within a few weeks — from the first audited batch change.
• Indicative payback between 4 and 9 months, depending on the batch change frequency, the current cost of the pre-audit sprint (quality hours + floor hours) and the cost of an eventual non-conformity or re-audit.
• The hard lever is not losing the Crossed Grain: a re-audit, the explanation to the retailer that requires it, the risk of shelf withdrawal for an indefensible "gluten-free" label. A single major non-conformity avoided pays for the pilot.

And the fair question from the quality manager

"What if the AI assembles a dossier that looks correct but is not?" — the evidence pack iLEAN generates is an anchored task, not free generation: the agent takes real photos Edge captured, MES data and the operator's human signature, and joins them in a template. It invents nothing. In anchored tasks, the best models brought the error below 1.5% ^[1]. And even then, nothing critical is decided alone: the person signs on the spot before the dossier closes. The three safety rings are there precisely for this.

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