Edge blocks startup until mold and color are validated

The problem

The riskiest moment in a closure plant is signed off by eye.

A mold or color change is the moment of highest risk: a wrongly mounted cavity or residual color can ruin the first tens of thousands of pieces. Verifying that 'mold and color are correct' today depends on a technician confirming it visually, with no structured evidence.

• A mold or color change is the moment of highest risk: a wrongly mounted cavity insert, a blocked cavity left open by mistake or residual masterbatch still in the barrel.
• At closure cadence, a bad startup can ruin the first tens of thousands of pieces before anyone notices the shade is off or one cavity is short. Those closures still have to be sorted, reground or scrapped.
• Today, "mold and color are correct" depends on a technician confirming it visually, under time pressure, with no structured evidence behind the signature.
• When a bottler finds mixed shades in a batch, nobody can show what the press looked like at startup. The changeover left a tick on a sheet and nothing else.

How it fits the IRIS system

Edge with JIDOKA AI and SMED AI — no full speed until every camera says OK.

ILEAN Edge places fixed cameras at the press's critical points that compare the current state against the active order's reference. It won't let the press start at full speed until every camera says OK and the quality supervisor signs off with visual evidence. SMED AI speeds up validation so it doesn't penalize changeover time.

The cameras do not replace the quality supervisor: they give her something to sign on. Each critical point is compared with the active order's reference, the press only goes to full speed when everything matches, and SMED AI keeps the validation from adding time to the changeover. Production gets its full speed; quality gets its evidence; nobody waits for the other. The changeover ends with proof, not with a hope that the first bags are right.

See the full IRIS architecture →

Before and after

Mold and color changeover today versus changeover on evidence

Aspect	Today	With iLEAN Edge
What "correct" is based on	A technician's look	Images against the order reference
Cavity inserts and blocked cavities	Assumed	Checked by camera
Residual color after purge	Judged by eye	Compared with the order shade
Mismatch at closure startup	Found in the bags	JIDOKA AI blocks full speed
Time added to the mold change	—	Not penalized: SMED AI
Record of the mold change	A tick on a sheet	Before/after images and signature

Eyeballed verification on every changeover → startup based on visual evidence, with automatic blocking.

Impact estimate

Impact estimate — tied to how often you change mold and color.

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.

• We do not fix a payback range here, because it depends on your mold and color changeover frequency; we calculate it with your own changeover count.
• Less startup scrap after every changeover: the first shots at full speed are made with the right mold, cavities and shade. Fewer closures to sort, regrind or scrap after every change.
• Protection against color cross-contamination, which is what turns a single bad startup into a bottler's batch rejection. Mixed shades in a bag are among the complaints bottlers tolerate least.
• And every changeover leaves before/after images and a signature that the audit dossier can use later without extra work from quality. The changeover documents itself while it happens.

Payback depends on changeover frequency · cuts startup scrap. Reduced startup scrap after every changeover and protection against color cross-contamination. Payback depends on mold/color changeover frequency. Estimate to validate.

And the fair question from the production manager

“Won't the camera hold up a changeover for nothing, with the bottler's order waiting?” — it stops on a verifiable mismatch against the order's reference, not on uncertainty. Comparing a known mold and shade with its reference image is an anchored task, where the best models drop below 1.5% error [1]. When the doubt is the model's, it escalates to the quality supervisor instead of blocking silently. The supervisor sees the two images side by side and decides in seconds. Nobody is left wondering why the press will not go to full speed. The reason is on the screen, with the image that triggered it.

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