Validate the formula in seconds, no typing

The problem

The most sensitive data in the plant enters without anyone signing it.

a formula or allergen error isn't caught today until the customer complains, weeks later.

• The exact composition of a blend and the allergens it declares are the two data points that can force a recall. Everything else in the plant is a quality question; these two are a legal one, and they are handled today with the same clipboard as everything else.
• Today they are written on the formula sheet and typed later, often by somebody who was not at the mixer and did not see which sack went in or which version of the recipe was on the clipboard.
• A wrong allergen on a label is not a complaint, it is a product withdrawal, and it surfaces weeks later when the customer's own audit catches it, by which time the product is on a supermarket shelf.
• When the certification body asks who approved that formula version and on what basis, the answer has to be a name and a timestamp, not a checkbox in a spreadsheet that anyone could have ticked.

How it fits the IRIS system

Connect with early human verification — two taps at the mixer.

industrial tablet with a visual summary of what was captured (photo, voice or spreadsheet); confirmation in seconds, no retyping.

It is the piece that guarantees the central memory is not contaminated and that every formula has a responsible person by name. Without it the other eleven do not get approved: nobody signs off on AI writing an allergen declaration unsupervised, and they are right not to. The gate is what makes the rest of the architecture defensible.

See the full IRIS architecture →

Before and after

Today's formula entry versus the signed formula

Aspect	Today	With iLEAN
When the data enters	Hours later, by transcription	In seconds, at the mixer
Who appears as the author	Whoever typed it	Whoever was at the mixer
What the sign-off proves	That somebody signed	Which formula and allergens they saw
Allergen error	Found when the customer calls	Caught before the batch runs
Manager's time	Walk to a terminal	Two taps on the tablet
Who writes to the system	A person, typing	A person, confirming

unverified data entering blind → a seconds-long human gate before ERP/MES.

Impact estimate

Estimated impact — to validate 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.

• No standalone payback: this is an enabling piece. Its value is risk, not productivity, and it is presented as such.
• What it removes is the undetected formula or allergen error, whose cost is not a line on a budget but a product withdrawal and the commercial relationship that goes with it.
• It is also the gate that makes the rest approvable, because it is what keeps unverified data out of the central memory that every other case writes into.
• And every correction the manager makes is stored with its author: the system learns exactly where its extraction fails on your formulas, and the audit trail shows who caught what.

reduction in undetected formula/allergen errors. *Estimate to validate*.

And the fair question from the production manager

"If a person signs at the end, what is the AI actually saving?" — the work is not in deciding, it is in assembling. Today somebody would have to find the batch, read the sheet and type the components. Here it arrives assembled and they confirm or correct. Extraction itself is an anchored task where the best models drop below 1.5% error [1], and nothing crosses without the sign-off.

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