Here is an uncomfortable observation: the easiest audit to pass is one written for a product the auditor has seen a thousand times. Tesla is pushing Optimus toward mass production, and pre-production audits of suppliers across the Yangtze delta are running right now – against a checklist borrowed from the car industry, for a machine nobody has ever mass-produced. I have spent two decades running supplier audits under VDA 6.3 and IATF 16949, automotive and aerospace. I know what those audits certify. I know what they cannot.

What a pre-production audit can actually verify

Strip an audit down and four load-bearing parts survive. A PFMEA where severity and occurrence are argued with data, not folklore. Control plans whose checks map to real failure modes instead of whatever the last customer asked for. Reaction plans with named owners and clocks on them. Run-at-rate evidence that shift three performs like shift one. That is process maturity, and it is worth certifying.

What no audit certifies is reliability, because reliability is a property the field grants, not a document. Every actuator, tendon route and tactile skin on that robot is a design in motion, meeting production for the first time. IATF 16949 has many virtues. Prophecy is not among them.

Process discipline can be audited in two days. Product reliability can only be earned in the field.

The borrowed checklist

The automotive acceptance stack – dimensional Cpk, PPAP elements, torque verification – is being mapped onto humanoid failure modes, and the seams show. Cpk measures deviation from nominal at time zero. The failure mode of a robot joint is drift: backlash creeping into an actuator over twenty thousand duty cycles, gear and tendon fatigue, lubricant migrating through a supposedly sealed joint, tactile sensors degrading after a million grasps, a dense battery pack managing heat in a chassis that moves like a body instead of rolling like a car.

A bolt at torque spec tells you nothing about the harmonic drive behind it. A PPAP file assumes a frozen design – the one thing an unprecedented product does not have. And there is no IATF 16949 for humanoids. The buyer writes the standard, then grades its own exam, which means the audit measures the buyer's understanding of the product as much as the supplier's. When both sides are discovering the failure modes together, the checklist had better target the physics you actually fear, not the physics your templates remember.

Audit the learning, not the ledger

With no track record, the only honest proxy for future performance is learning velocity – how fast an organisation turns a surprise into a countermeasure. That is observable. It is also measurable.

  • QRQC habit. Does the shop floor stop at the defective part and solve within 24 hours, or does every problem funnel into a weekly quality meeting? Ask to watch one happen.
  • Change-order latency. Pull the last ten engineering changes. Count the days from detection to PFMEA revision. A supplier that updates its FMEA within days of a new failure mode is worth more than a wall of certificates.
  • The last novel failure. Ask for the 8D on the strangest thing that ever hit their line. Closure time, recurrence rate, whether the lesson survived staff turnover.

I lived this from the other side. At SNOP I built a QA/QC department from zero for a 900-plus-employee greenfield plant – no PPM history, no warranty curve, nothing to audit but the machine we were becoming. You cannot audit what has not happened, so we built detection speed to stand in for a track record. Defect costs fell 70%; customer satisfaction reached 98%. Earlier, at WITTE Automotive, QRQC and A3 drove a substantial failure-cost reduction – closure times shortened, recurrence dropped. Learning velocity is not a vibe. It shows up in the numbers.

Give me a Yangtze delta supplier with three ugly, closed 8Ds and a worn QRQC board over one with a spotless audit file and no scar tissue. The first can be trusted with an unprecedented product. The second has simply not been tested yet.

Who finds the first failure at rate

Every accelerated launch answers one question: who detects the first failure at full rate – the supplier, or you. Acceleration does not remove physics. It moves the discovery of failure downstream, into volumes where one weak joint becomes a fleet problem, unless detection is bought up front. Safe launch means 100% containment on the first lots, telemetry tripwires through early production, exit criteria per lot, and escalation triggers at defined defect rates – not "we'll watch it." A containment team and a full sort will run you €80,000 a week. A humanoid failing in a customer's facility costs multiples of that, plus credibility you cannot buy back.

Key takeaways

  • A pre-production audit certifies process discipline – PFMEA honesty, control-plan linkage, reaction speed, run-at-rate repeatability. It never certifies reliability; the field has not happened yet.
  • Borrowed automotive criteria go quiet on humanoid physics: backlash drift, tendon and gear fatigue, lubricant migration, sensor degradation, battery thermal behaviour. Write the checklist for the failure modes you actually fear.
  • With no history, audit learning velocity: QRQC habit, change-order latency, days from novel failure to PFMEA revision. Closure time and recurrence rate are the track record.
  • Buy detection before you buy volume – containment on early lots, tripwires at rate, hard escalation triggers. Acceleration without detection only schedules the discovery for a worse week.

Somewhere in the Yangtze delta a quality manager is signing a torque record for a joint that will one day fail in a way torque records cannot see. No audit was ever going to catch it. The audit's honest job is narrower and more useful: establish whether the supplier learns faster than the failure modes arrive, and whether the tripwires are live before the rate goes up. Get that right, and an unprecedented product earns its own history instead of inheriting a borrowed one.