No line starts without evidence

The problem

The mistake takes five minutes to make and gets paid for over weeks.

If the purge is incomplete, the first parts come out with color or material carryover — an outright reject on visible parts. If the wrong mold gets mounted, or the blocked cavity was not the one it should have been, a whole batch gets produced that is not what the customer ordered. The mistake takes five minutes to make and gets paid for over weeks. Today it all rests on somebody signing that it is ready, under pressure to recover machine downtime and with no record of what was actually verified.

• Incomplete purge: the first parts come out with color or material carryover, an outright reject on visible parts.
• Wrong mold, or the wrong cavity blocked: a whole batch gets produced that is not what the customer ordered.
• Today it all rests on somebody signing that it is ready, under pressure to recover machine downtime.
• And there is no record of what was verified, so afterwards there is nothing to review.

How it fits the IRIS system

Edge plus JIDOKA AI — startup blocked until every point reads OK.

Edge plus JIDOKA AI.

The decision is still a person's. The AI supplies the evidence to sign on, not the signature.

• Cameras at the critical points — mounted and identified mold, nozzle and screw after purging, material and color in the hopper, robot and part takeout — trained to recognize the “ready to start” state.
• Startup stays blocked until every point reads OK.
• The lead signs on that evidence, point by point.
• The decision is still a person's: the AI supplies the evidence to sign on, not the signature.

See the full IRIS architecture →

Before and after

Blind changeover vs. changeover with evidence

Aspect	Current changeover	With Edge and JIDOKA AI
Purge check	By eye, under time pressure	Camera on nozzle and screw
Mounted mold	Assumed correct	Identified and verified
Material and color	Checked at the hopper if there is time	Verified point by point
Startup	Whenever the machine is ready	Blocked until everything reads OK
Signature	Blind	On visual evidence
Carryover	Detected by the customer	Before the first good part

From a blind signature under time pressure to a signature on point-by-point visual evidence; from color carryover found by the customer to found before the first good part.

Impact estimate

Impact estimate — to validate against your 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.

• Several mold and color changes per day, serving markets with different rules.
• Indicative payback between 6 and 12 months depending on the number of changeovers per day.
• Between reduced startup scrap and, above all, avoiding the contaminated batch that triggers a containment.
• A single avoided containment usually covers a good part of the deployment.

Estimated payback 6 to 12 months depending on the number of changeovers per day, between reduced startup scrap and, above all, avoiding the contaminated batch that triggers a containment. Estimate to validate.

And the fair question from the production manager

“Won't blocking startup cost me more than it saves?” — it is a fair objection, and it is why this case travels with SMED AI: the goal is a changeover that is faster and verified, not one that is safer but slower. The check adds seconds at points that were already being looked at; what it removes is the whole batch produced with the wrong mold, which is hours of machine time plus the containment.

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