Here is an uncomfortable observation from two decades of qualifying suppliers with a launch clock running: when the buyer's own engineers walk a supplier's plant, the audit stops measuring capability and starts measuring willingness. Reports from the Yangtze Delta describe Tesla's robotics team touring supplier factories ahead of the Optimus mass-production push, with orders landing within days of the visits. That is not a process audit. That is sourcing speed-dating run by the buying side. I ran VDA 6.3 supplier process audits at Witte Automotive, then a full qualification sprint while building the QA/QC department of a 900-person greenfield plant at SNOP. I know what a plant looks like when it knows visitors are coming, and what it looks like at two in the morning when nobody is watching. Two different factories. Only one of them ships your parts.

An engineering walk-through is not a process audit

VDA 6.3 survives as a discipline because of two design decisions: the auditor stands independent of the purchasing decision, and findings are triangulated – documents, floor observation, interviews at the machine, capability records. Remove either and you have a tour with a scoring sheet. An engineering walk-through answers one question: can these people make this thing once, beautifully? A process audit answers a different one: will this process make the same part on the third shift in week forty, after two operators have left and the maintenance tech is on holiday.

At Witte I watched a forming supplier charm a customer visit. Spotless lines, briefed operators, a trophy shelf of parts machined to nominal. When we returned unannounced for run-at-rate, the third shift was a different country. The die swapped out for the tour was back in the press. Cpk on the pivot bore read 0.8 against the 1.67 the tour samples had shown. The rework bench had grown a containment shelf nobody had mentioned. Six weeks of re-qualification cost us perhaps €80,000 in engineering hours – a fraction of the field escape we would otherwise have owned.

A supplier audit run on the buyer's clock measures the plant's best day; run-at-rate measures every other day.

The failure library is empty

Humanoid robots have no field history. Every line of an automotive checklist – fastener audits, weld spatter, torque-to-yield bolts – exists because some plant somewhere shipped garbage on a Tuesday and somebody wrote it down. A checklist is yesterday's failure library, bound and laminated. Point it at a new product category and it goes politely blind: harmonic-reducer wear after ten thousand hours of reversed cyclical load, cable fatigue inside a forearm, thermal drift in actuator windings, backlash decaying into hysteresis. No recycled VDA question catches any of that. No plant has bled for it yet.

The honest substitutes are stress-to-failure testing, worst-case boundary builds and durability run on the real duty cycle – not the spec-sheet one. A PFMEA for a new category should start nearly empty and be earned line by line in the lab. An inherited one is someone else's homework, answers filled in, for the wrong exam. I design and deploy autonomous multi-agent systems in my own work, so this is not an argument against new machines. It is an argument against pretending a new machine carries an old machine's scars.

When the audit becomes a purchasing gate

Audit-to-order compression changes what the audit optimises. When orders land within days, the rewarded skill is staging: the best-behaved plant wins, not the most stable process. And when the team that needs parts next month writes the findings, red becomes amber by Thursday. Independence is not bureaucracy. It is the only reason a score means anything.

The artefacts that predict escapes never appear on a tour. Capability data across all shifts. Escape history – the last 8D reports and what permanently changed because of them. Change-control records. That last one is decisive. In years of de-escalation work, cleaning up customer complaints born in processes nobody had vetted, most escapes traced back to a change someone made without validation – a moved datum, a quieter sub-supplier, a cycle time trimmed by two seconds. The line had been qualified once. The process underneath it had moved.

So I skip the trophy shelf and ask for three things:

  • Capability data on the characteristics that matter to my part – last three months, all shifts, not the golden-sample drawer.
  • The last five process changes and the validation that released each one.
  • The last 8D reports and what permanently changed because of them. A supplier with no 8Ds is either flawless or blind, and I have yet to meet the flawless one.

Key takeaways

  • An audit run by the buying team under launch pressure measures willingness, not capability – independence is what gives findings meaning.
  • A product category with no field history needs stress-to-failure testing, worst-case boundary builds and real duty-cycle durability; treat every inherited checklist as provisional.
  • Capability data across shifts, escape history and change-control logs predict escapes better than any plant tour.
  • Most escapes come from unvalidated process changes – audit the change process harder than the line itself.

The industry checklist for humanoid robots will be written exactly once – either in a lab, by engineers who ran components to failure before the field did it for them, or from field escapes, by whoever inherits the warranty. Somebody's audit signature precedes every line of it. If you are that somebody, sign for data you generated, not for a tour you were given. The robot has no history. Run-at-rate, capability and change control are how you give it one – before the field writes it for you.