Instrument cluster control with AI — the end-of-line test samples seconds; the fine cosmetic defect reaches the OEM.

The problem

The cluster passes the functional test — and the OEM finds the speck under the glass.

The quality manager on a cluster line tells the same story in every plant: the panel passed the end-of-line test, communications OK, actuators OK, the display sample the test plays back looked fine. And weeks later the OEM claim arrives: a speck under the glass, a brightness gradient you only see driving at night, an icon that lights up in a shifted color. The defect did not fail a single electrical parameter — it was cosmetic, fine, and it got through.

Three things happen at once on an instrument cluster line, and almost nobody covers them together:

1. The functional test samples, it does not verify — it plays back a few seconds of a display sequence that runs for minutes. Dead pixels outside the sample, brightness uniformity, the icon that must light up exactly right in each variant are left without 100% verification.
2. The fine cosmetic defect gets past the operator at takt — the speck under the glass, the uneven brightness, the clip that only half closed. The human eye at takt cannot sustain it shift after shift; the veteran catches it more often, but not always.
3. The firmware variant is checked by hand — market, language, option package. A cluster flashed with another market's variant is functionally perfect and commercially useless, and it is discovered late.

The result is always the same: the OEM finds the escape, the plant pays for the claim and the sorting, and the discussion with the display supplier — did the defect come in the display or did assembly create it? — stays at the level of opinions because there is no evidence per batch.

How it fits the IRIS system

iLEAN does not replace the end-of-line test — it covers what the test cannot see.

The cluster problem is not a lack of testing: the line already has a functional test, serial number traceability and a flashing station. The problem is that each control covers its own island — the test samples the display, the operator looks at takt, the variant is matched by hand. iLEAN acts as the putty that binds together 100% verification of the sequence, the closing of every clip, the OEM order and the history per display batch, without asking you to change the test or the flashing station.

Edge verifies the full sequence pixel by pixel and the closing of every clip. Connect matches serial number, OEM order and firmware before flashing. JIDOKA AI holds on drift. The person decides.

The four iLEAN pieces applied to instrument cluster control:

• Edge — local vision over the end-of-line station: it captures the full cluster power-up sequence and compares it pixel by pixel against the reference pattern for each variant — dead pixels, brightness uniformity, every icon in its exact color and position — in seconds, on 100% of the parts. At the assembly station it cross-references acoustic signature and image to verify that every clip actually closes, not that it looks closed. It runs locally: if the plant loses the network, Edge keeps capturing, comparing and holding. What is critical does not depend on WiFi.
• Connect — stitches the cluster serial number to the OEM production order (market, language, options) and to the firmware variant the station is about to flash. If they do not match, nothing is flashed. Matching stops being a manual check and becomes an upfront block with a record per serial number.
• JIDOKA AI — if the defect rate drifts (same defect repeated, same area of the glass, same shift), it holds the flow and raises the alarm before the batch ships to the OEM. It never acts alone on anything critical: it proposes the hold, the line manager decides.
• Agents — they correlate the cosmetic defect detected by Edge with the display supplier's batch, the fixture, the shift and the variant. When the speck or the brightness gradient concentrates in one received batch, the agent assembles the evidence — images, serial numbers, dates — for the claim to the supplier.

See the full IRIS architecture →

Before and after

Classic cluster control vs. control with iLEAN

Aspect	Functional test + inspection at takt	With iLEAN Edge + Connect
Display sequence	A few seconds sampled in the test	100% of the sequence, pixel by pixel
Fine cosmetic defect (speck, uneven brightness)	The operator's eye at takt	Vision against a pattern, in seconds
Clip closing	Visual inspection and feel	Acoustic signature + image for every closure
Firmware/language variant per market	Manual check against the order	Serial number matched to the OEM order before flashing
Defect rate drift	Discovered in the container or at the OEM	JIDOKA AI holds the flow and raises the alarm
Claim to the display supplier	No evidence per batch, opinions	Defect correlated with the display batch, with images

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Instrument panel/cluster line with an end-of-line functional test, visual inspection at takt, a station that flashes market variants, and OEM claims for cosmetic defects documented over the past months.
• Edge pilot over the end-of-line station + Connect stitched to the OEM order. First value expected within a few weeks: 100% verification of the sequence and variant blocking work from day one, before JIDOKA AI and the Agents fine-tune with the history.
• Expected reduction of cosmetic defect escapes to the OEM of ≥30% in the first months — estimate to be validated with your claims data.
• Target of zero wrong firmware variants with serial number ↔ OEM order matching before flashing, with a record per serial number for audit.
• Indicative payback between 5 and 12 months, estimate to be validated. The hard lever: every cluster that does not reach the OEM defective avoids a claim, sorting and emergency freight — the three costs that hurt most in PPM.
• Recurring benefit that does not go into the ROI but carries weight: the evidence-backed claim to the display supplier — correlated images, serial numbers and batches — recovers cost that today stays with the plant for lack of data.

And the quality manager's reasonable doubt

“What if vision flags false defects and holds good parts at takt?” — the iLEAN system never scraps on its own. Edge compares against a pattern and proposes; JIDOKA AI holds on drift; the person confirms or releases. Hallucination is a problem of free generation, not of anchored tasks: in tasks where the AI compares a captured sequence against a reference pattern and a history, the best models brought error below 1.5%^[1]. And even so, what is critical goes to the safety rings — the agents live in the outer ring, they propose inward, and holding or releasing a part is signed by a person. Never the other way round.

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