For seventy years we demanded a calibration certificate from a €40 caliper. A sticker, a due date, an unbroken trail back to the national standard – and an auditor who will find the one expired gauge in a cabinet of four hundred without breaking stride. Then we installed six-figure vision systems whose software updates itself overnight, and nobody asked for a signature. I have stood on both sides of that fault line. Two decades of IATF 16949, AS9100 and VDA 6.3 audits across automotive and aerospace, including standing up the QA/QC department of a 900+ employee greenfield plant from a bare hall. And, in parallel, the software-supply-chain habits of a certified ethical hacker. The gap between those two worlds is where the next generation of quality failures is already growing.

When the gauge ships with a changelog

Metrology and Quality News ran two pieces this week that belong together more than their authors intended. One calls for a version passport for every AI measurement; the other charts optical inline quality assurance spreading through electronics plants. Side by side, they describe an industry bolting neural networks onto lines whose traceability apparatus is still calibrated for steel.

When a neural network replaces the caliper, a measurement stops being one number from one instrument. It becomes the output of a stack: model build, threshold configuration, image preprocessor, camera firmware. Nudge a confidence threshold by 0.02 and you will move more accept/reject verdicts in one shift than a hardened gauge block drifts in a year. Push a firmware update that quietly changes auto-exposure behaviour and you have shifted the input distribution of every part measured since – the model has no idea, and neither will you. The caliper never phoned home at three in the morning to install a patch.

My answer comes from security, because this is a supply-chain problem wearing a lab coat. Nobody sane runs unsigned builds in production software. A vision system deserves the same discipline: signed model builds, a software bill of materials for the whole stack, and a hash on everything that can change a verdict.

The audit question nobody can answer yet

Here is the question that lands sooner than anyone is ready: “Which model version accepted lot 4711 on the night shift?” “The system was up to date” is not a traceable answer, and no certification body will accept it as one. The logic already sits in the standards – AS9100 drags software into configuration management outright, and IATF 16949 expects revision control when measurement programs change after an engineering change. The muscle memory is not on the shop floor.

I paid for the analog version of this lesson at WITTE Automotive. A customer's incoming inspection rejected a latch dimension we had certified good; their CMM and ours disagreed by 40 µm. The QRQC session found the root cause in ninety minutes: three weeks earlier, a technician had improved our CMM program – corrected probe path, better alignment, genuinely better – but nothing recorded which program revision had produced the certificates for lots already on a truck. We could not prove the shipped parts were measured under the old program. So we could not defend the results. Three lots quarantined, 21,000 parts sorted by hand across a weekend, €24,000 gone in sorting and premium freight, plus an 8D report and the kind of customer silence that costs more later. That was one measurement program with no version control. A modern line runs the same failure with a model that retrains itself.

A measurement you cannot attribute to a version is not a result – it is a rumour with a decimal point.

A version passport on a running line

The passport is not exotic. It is a calibration certificate rebuilt for software. Every verdict record is written at the moment of measurement and carries the model identifier and build hash, the training-data tag, the threshold configuration hash, the preprocessor revision, the camera firmware, the line release. The reference image set is frozen per model version – golden panels, exactly like golden parts, except made of pixels. A model update is treated as a gauge change, because that is what it is: pinned to a line release, requalified against the golden set, MSA refreshed, signed off. No silent updates on a production line. Ever.

I built the same discipline into my own AI work. MultiPS, the multi-model orchestration platform I designed, runs 63+ models in parallel, and every output carries its model identity and version before synthesis. A consensus number without provenance is decoration – whether it comes from a committee of models or one camera over a conveyor.

Decide now how you would scope a recall. Optical inline QA is rolling out across electronics plants, and the EV sector generates recall headlines fast enough to keep every buyer twitchy; a bad part will eventually escape. The question is whether “which lots did version 4.2 pass” answers in minutes – or whether you re-inspect everything the system has ever touched.

Key takeaways

  • Treat every model update as a gauge change: pin model build, threshold config and camera firmware to the line release, then requalify against a frozen golden-panel set before the first production part runs.
  • Write model identity, build hash and config hash into each verdict record at the moment of measurement – not into a vendor's server log you will discover during the audit.
  • Disable auto-updates on production vision systems; accept signed builds only, and keep a software bill of materials for the full stack, cameras included.
  • Build the recall query before you need it: “which lots did version X pass” must answer in minutes, because the alternative is re-inspecting everything the system has ever touched.

The next auditor will not ask whether your AI inspection is accurate. Accuracy is assumed, modelled, demonstrated on slides. They will ask which version said so – and you will answer with a query in minutes, or with quarantine tape and re-inspection. We spent seventy years earning the right to say trust the certificate. The certificate is now a changelog. Sign it.