The most expensive planned stop restarts today on a signature with no evidence

The problem

The most expensive stop of the week restarts on a signature with no evidence.

Changeover concentrates the whole week's risk into a window of a few hours. A scrap offcut left in the die destroys the tooling on the first stroke. A set mounted with the wrong code produces a part number that is not the one on the order. Centring outside tolerance produces a whole lot with runout out of specification that is only discovered at final inspection, once the entire finishing process has already been invested in it. And the pressure to recover the lost hour always pushes the same way: start before you are entirely sure. This is not a discipline problem, it is an incentive problem. Exhaustive checklist verification, the classic answer, collides with that same pressure: it lengthens the stop, so it gets filled in afterwards.

• The tooling changeover concentrates the whole week's risk into a window of a few hours: one to three, to swap the die set and readjust the roll former rollers.
• A scrap offcut left on the die face destroys the tooling on the first stroke — tens of thousands of euros per incident.
• A die set mounted under the wrong code produces a part number that is not the one on the order.
• Centering out of tolerance produces a whole lot with out-of-spec runout, and it only surfaces at final inspection, once all the finishing has been invested in it.
• Pressure to recover the lost hour always pushes toward starting before you are entirely sure. It is not a discipline problem, it is an incentives problem — and that is why the checklist gets filled in afterwards.

How it fits the IRIS system

Edge with JIDOKA AI, plus SMED AI pulling the other way.

Edge with JIDOKA AI, plus SMED AI.

JIDOKA AI and SMED AI offset each other on purpose: one prevents starting too early and the other shortens the window. Without the second, the first would be experienced as a brake; together, the changeover comes out shorter and correct.

• Edge cameras at the N critical points of the changeover.
• Comparison of the current state against the mounted-and-ready reference state for that specific part number: tooling code visible and correct, position and centring within tolerance, die clean and free of offcuts, lubrication done, no tool left inside.
• JIDOKA AI: until the cameras confirm and the supervisor signs on that evidence, the press does not start. The line stops itself before producing badly, not after.
• SMED AI works the other way: it guides the changeover sequence step by step, orders the tasks that can overlap and learns from the history of changeovers on that same part number to shorten the window. Verification cannot cost productivity, or it will not be adopted.

See the full IRIS architecture →

Before and after

Today's changeover versus the validated changeover

Aspect	Today	With iLEAN Edge
What authorizes start-up	A signature with no evidence	Camera conformity, then a signature on it
Offcut left on the die face	Found on the first stroke	Detected before start-up
Wrong tooling code	Found at final inspection	Detected before start-up
Centering out of tolerance	A whole lot with runout out of spec	Blocked before producing
Changeover duration	1-3 h	Guided step by step and shortened
When the line stops	After producing badly	Before producing badly

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.

• Estimated payback 5-10 months, and probably the easiest to calculate with your own data.
• Availability hours recovered at the plant's bottleneck.
• Tooling damage avoided: tens of thousands of € per incident.
• Out-of-tolerance lots that never get produced. Two of your numbers are enough: how many changeovers per week and how long they average.

estimated payback 5-10 months, and probably the easiest to calculate with your own data: availability hours recovered at the bottleneck, tooling damage avoided (tens of thousands of € per incident) and out-of-tolerance lots never produced. Two numbers are enough: how many changeovers per week and how long they take on average. *Estimate to validate.*

And the fair question from the production manager

«What if the camera blocks the line by mistake and stops my production?» — that is the right objection, because a false stop at the bottleneck costs real money. This is why conformity is checked against the reference state for that specific part number and, in case of doubt, the system does not block silently: it escalates to the supervisor, who signs on the evidence. The line stops on a verifiable discrepancy, not on model uncertainty.

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