The new color does not start until the evidence says it can
The color changeover is the most delicate moment in a paintshop: if the start goes wrong, half an hour of production comes out contaminated — parts already painted and cured, the most expensive scrap there is. Today that depends on a signature given under sequence pressure and with no data. With iLEAN Edge, cameras at the critical points compare against the reference state and the line does not start until everything clears and quality signs, with the evidence in view.
A signature given under sequence pressure decides whether half an hour comes out good or goes to scrap.
Every color changeover ends with the same phrase: "ready to start". That signature is given by someone who knows what they are doing, but who gives it with the sequence pressing and no objective evidence in front of them. And the price of being wrong is not proportional to the gesture:
- Startup scrap arrives in bulk — it is not one part that fails, it is the whole run until somebody notices. And these are parts already painted and cured, meaning the most expensive scrap in the process, not raw material.
- The signature rests on nothing verifiable — you sign that the circuit is purged and the booth is clean, but no record remains of what state they were actually in. When the failure has to be analyzed afterwards, there is no data to go back to.
- Sequence pressure always pushes the same way — when the day's plan is tight, the temptation to start "it is fine already" is structural. It is not a discipline problem, it is the problem of having no cheap way to check.
The result is recurring color startup scrap that everyone accepts as part of the process, when it is really the consequence of signing blind.
Edge + JIDOKA AI — clean vs. residue per critical point, and a block until every one clears.
Here the AI does not replace anyone's judgment: it puts evidence in front of it and removes the pressure of deciding without data. Cameras at the critical changeover points compare the real state against the reference state and return a verdict per point, in seconds.
Purge circuit, booth, bells and first part: each point has its camera and its clean-vs-residue classification. The line does not start the new color until all of them clear and quality signs — this time with the evidence in view, not from memory.
How Edge works on the color changeover:
- One verdict per point, not one overall — the circuit can be purged and the booth not. Each critical point is assessed separately, so if something fails you know exactly what to repeat.
- JIDOKA AI: block, not alert — while any point is red, the start is not enabled. It is the jidoka principle applied with vision: the process stops itself before generating defects, rather than reporting them afterwards.
- The signature remains human — the system does not start the line on its own: it enables the start and quality signs. What changes is not who decides, but what information they decide with.
- The complete event recorded — photos of every point, owner, timestamp and result are stored as a single event. That record later becomes the color changeover evidence for audit, with no additional work.
- Trained on your colors — residue of a red on a white looks nothing like residue of a grey on a black. The model learns from the real changeovers in your catalog, which are the ones that will actually happen.
Color changeover signed blind vs. validated with evidence
| Aspect | Signature without evidence | With iLEAN Edge + JIDOKA AI |
|---|---|---|
| Basis for the decision to start | Judgment under sequence pressure | Visual verdict per critical point |
| Points checked | Whatever there is time to look at | Circuit, booth, bells and first part |
| Starting with a point still red | Possible | Blocked until it clears |
| Color startup scrap | Recurring and in bulk | Drastically reduced |
| Record of the event | The signature, unsupported | Photos, owner, timestamp and result |
| Evidence for audit | Reconstructed afterwards | Already generated during the changeover |
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.
- Automotive paintshop with several color changes per shift and startup scrap accepted as a structural cost of the process.
- Edge pilot on the critical changeover points, with start blocking and event recording. Without changing the changeover procedure or who signs. First value expected within a few weeks.
- Indicative payback of 4 to 9 months, strongly dependent on two variables of yours: how many color changes you run per day and what a contaminated start costs. Estimate to be validated.
- The main lever is avoiding scrap in bulk: you do not save one part, you save half an hour of already painted and cured production every time a bad start is prevented.
- A derived benefit that appears on its own: every color changeover is documented with photos and a signature, so the evidence auditors ask for about color changeovers stops being prepared after the fact.
And the fair question from the production manager
"Is this not going to stop my line every five minutes?" — the block only acts when a critical point is red, which is exactly the case where starting is expensive. On the reliability of the verdict: classifying an image against a pattern trained on your own colors is an anchored task, not free generation, and there the best models brought the error below 1.5% [1]. And the last word still belongs to quality: the system enables the start, the person signs it.
[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.
What people ask about validating the color changeover with vision
Which points are checked exactly during the color changeover?
The ones that determine whether the new color starts clean: the purge circuit, the booth, the bells and the first part painted in the new color. Each has its own camera and its own verdict, because grouping everything into a single "OK" loses the information that matters: if something fails, you need to know which specific point to repeat. The list is adjusted during the assessment to the plant's real circuit, because not every paintshop has the same purge configuration or the same points where residue accumulates.
Does blocking the start penalize line capacity?
It only blocks when a point is red, and that is precisely the start you did not want to make. In practice the balance is very favorable: the check costs seconds and prevents half an hour of contaminated production, which is also the most expensive kind of scrap — parts already painted and cured. On top of that, verification reduces blind repetitions: when the operator knows which point is wrong, they repeat that one and not the whole procedure. If the plant runs many changeovers per shift, that combination of SMED and verification is what stops verifying from costing capacity.
Who finally decides to start, the system or the person?
The person, always. The system enables the start when every point clears, but the quality signature still exists and is still what authorizes. What changes is not who decides but what they decide with: today the signature is given under sequence pressure and without objective data; with Edge it is given with the visual evidence of every point in view. It is an important difference on the human side too, because it takes off the responsible person the burden of carrying alone a risk they had no cheap way to check.
How does it tell residue apart from a shadow or a reflection?
Because the model is trained on the real changeovers in your color catalog, not on generic images. Red residue on white and grey residue on black are completely different visual problems, and so is the lighting at each point of the installation. By learning from your own changeovers — good cases and bad cases — the CNN learns where the boundary lies in your installation. And as in the other Edge cases, discrepancies a person corrects feed back into the model, so the margin sharpens over the first weeks.
Does the changeover record count as audit evidence?
Yes, and it is usually the benefit nobody counted at the start. Every changeover is stored as a complete event: photos of each critical point, owner, timestamp and result. When the customer audit or the IATF asks for evidence on color changeover control, that evidence already exists, dated and signed, instead of having to be reconstructed from shift reports. It is the same information that later feeds the evidence pack, without anyone having to prepare it twice.
Tell us how many color changes you run per shift and we will send the number back.
We work on your plant's real data, not ours. With your changeover frequency and your contaminated-start cost we calculate the case. Assessment with no commitment.
See how we apply it in your plant — calculated on your own changeovers ‹ See all 12 paint & assembly cases See automotive