Here is an uncomfortable observation: your next supplier decision is one Google search away from being made by a content farm. A ranked comparison of five Chinese busbar suppliers for EV battery packs made the industry feeds this week — five names, tidy verdicts, the confident tone of a scoreboard. I have spent two decades qualifying suppliers under VDA 6.3 and IATF 16949 in automotive and aerospace, and I can tell you what that list cannot see: the part where the factory burns. A busbar is not a commodity bracket. It carries current inside a battery pack, and a bad weld or thin plating is not a cosmetic defect — it is a thermal-event path straight into the pack. No star rating on earth measures that.
What a ranking actually measures
Rankings measure what can be scraped from a website. Certifications, capacity claims, unit price, marketing polish. All useful. None of it is qualification.
What the list cannot see is whether the Cpk study was run on golden samples hand-picked for the customer visit. Whether the PFMEA in the PPAP file was written by an engineer who understands failure modes, or copied from a template with the part number changed. Whether the supplier answers an 8D in forty-eight hours or forty-eight days. And how they behave after an engineering change — because the supplier you sign is not the supplier you get three years and two revisions later, when a plating bath gets swapped to save €0.004 per part and nobody tells you until your customer's end-of-line test starts tripping.
I have audited plants that looked magnificent on paper and had no idea where their calibration certificates were. I have audited plants with peeling paint and a change-control process that would survive a regulator's interrogation without blinking. The ranking sees the paint. The audit sees the process.
What real qualification looks like on a busbar
For a busbar in an EV battery pack, qualification is unglamorous and non-negotiable:
- VDA 6.3 process audit — on-site, at the weld cell and the plating line, not a desk review of a PowerPoint.
- PPAP with plating-thickness capability — real studies on production tooling, real subgroup data, real measurement system analysis on the coating gauge.
- Weld cross-sections — sectioned, etched, pore and penetration evaluated, on every weld station. A weld that looks fine can be a void with a future.
- Dimensional stack-up in the pack — the busbar does not live alone; its tolerance interacts with the module, the carrier and the fastening torque.
- Run-at-rate — at contracted volume, contracted shifts, with contracted people. Capability at fifty parts an hour is a rumour.
- Change-control clauses with teeth — notification windows, revalidation triggers, financial consequence. A clause without a penalty is a greeting card.
Count the cost honestly. An audit week — travel, two auditors, three days on site, report — runs a few thousand euros. One escaped defect on a current-carrying part means containment, sorting at the customer, possibly field action: two orders of magnitude worse, sometimes three. At SNOP I ran quality for a 900-plus-employee plant and went a full quarter with zero critical customer escalations — not because we ranked suppliers well, but because the ones we trusted had survived audits. The same discipline underpinned the supplier-quality work FOREAST did for clients like ArcelorMittal. Nobody has ever de-escalated a crisis by showing the customer a leaderboard.
The leaderboard builds the shortlist. Only the audit signs the contract.
The disease has jumped to AI selection
Now the same illusion ships for software. AWS added quality, speed and cost rankings to its model catalogue, so you can pick an AI model the way a content farm picks a busbar supplier. The mechanics are identical: someone else's tests, someone else's data, someone else's edge cases, rendered as a score you can act on without ever visiting the factory.
Your production line does not run on someone else's edge cases. I built MultiPS, a platform that orchestrates 63-plus AI models in parallel with consensus synthesis, and the clearest lesson from running that system is this: models adjacent on public benchmarks disagree violently on the inputs I actually care about — ambiguous defect descriptions, malformed measurement data, the dull, ugly text real plants produce. Consensus measured against my own failure modes is the only ranking that counts. Public leaderboards had exactly one job in my workflow: building the shortlist. Never once the decision.
So treat a model the way a competent quality organisation treats a supplier. Define the qualification: your data, your tolerance for error, your requalification triggers when the vendor ships an update. A model version change is an engineering change. If you have no change control on it, you have learned nothing from a decade of supplier quality.
Key takeaways
- Rankings measure visibility, not capability — certifications and marketing tell you who to visit, never who to sign.
- For safety-critical parts, qualification is physical: process audits, capability studies on production tooling, weld cross-sections, run-at-rate, change-control clauses with financial teeth.
- Price due diligence as insurance: an audit week costs thousands; one escaped defect on a current-carrying part costs hundreds of thousands, plus the customer's memory.
- Apply the identical rule to AI: benchmark on your own data and failure modes, treat every model update as an engineering change, and let public rankings only ever build the shortlist.
The shortlist is cheap. The signature is expensive. Due diligence is boring on purpose — the boredom is the mechanism, the same way an audit checklist is boring right up until it catches the plating line running out of spec on third shift. The supplier you trust at 2 a.m., when the phone rings and containment is moving, is the one that survived the audit. It was never the one that topped the list.