The line is state of the art. The control limits are fiction.

Here is an uncomfortable observation: the most dangerous plant in any industry is the one where nothing has failed yet. This week produced two announcements of exactly that species – Statetron building an AI-and-robotics industrial facility in Qatar, and Daedong promising to turn its Daegu tractor works into a Physical AI hub that lifts production capacity by 67% by 2029. The headlines are all cobots and vision systems. Nobody asks the question that decides the first eighteen months: what, exactly, does the control chart plot in week one? A greenfield line has no history and no scars, and every number printed about it in month one measures your assumptions, not your process.

A new process has no priors

Control limits are not a property of the drawing. They are an empirical statement about what a process does when nobody interferes, and they want roughly 25 subgroups from a process that has stopped changing its mind. A new line never qualifies. Tooling wears in during the first weeks, fixtures settle into the floor, operators calibrate their hands. The limits you compute in month one are trial limits – provisional numbers that move with every shift until the process, not the statistics, decides where it lives.

Capability is worse. A Cpk study presupposes stability and a sane estimate of variation; a greenfield line offers neither, so the figure that lands in the customer's PPAP file is a prophecy, not a measurement. I have signed those figures, knowing they described a process three weeks old that had already been wrong twice – which made it twice as honest as a line still sitting at zero. Every control limit is statistical fiction until the process has been wrong a few hundred times.

Then there is the inspection AI, the part of these announcements that draws the applause. A vision model needs labelled images of every defect class – typically thousands per class – before it can be trusted to reject anything. A new process has not produced those defects yet. That is the point of it. The training data is a vacuum, and the model fills it with the only behaviour it knows: pass everything, with serene confidence. A system that has never seen a bad part is not an inspector. It is an expensive doorman.

A capability index is scar tissue expressed to two decimal places.

Build the boring parts first

The sequence that decides whether a greenfield plant survives is unglamorous. PFMEA and control plan get argued before the robots come out of their crates, because failure modes on paper cost nothing and failure modes discovered at rate cost everything. Then measurement systems analysis, before a single reading is trusted. If gauge R&R eats more than 10% of tolerance, your SPC chart is charting the gauge, not the part – and every study downstream inherits the error. Containment goes live on day one. Before there is a customer. Before there is a sellable part. The habit has to exist before the crisis does.

I lived this sequence. At SNOP I built the QA/QC department of a 900-plus employee plant from bare concrete – walls going up on one side of the fence, control plans being argued on the other. We signed the PFMEA and control plan before anyone was allowed to talk about line rate, and we ran measurement-system analysis until I trusted the gauges more than my own eyes. Only then did we chase output. Under that regime we cut defect costs by 70% and closed a quarter with zero critical customer escalations. Those numbers were not brilliance. They were the sequence, executed in order, by people who understood why the order mattered.

The temptation in every greenfield is to invert it – rate first, quality later, once we are stable. The arithmetic punishes you. I have watched a single escaped batch generate more sorting, premium freight and customer-imposed containment than a full year of MSA and safe-launch work would have cost. Sequence is the cheapest quality on the market.

Containment is how you buy time to learn

A safe launch is not an admission of weakness; it is a purchasing decision. You are buying time. 100% inspection on critical characteristics, enhanced surveillance on the rest, held in place until explicit exit criteria are met – a defined run of consecutive clean lots, a PPM ceiling, a capability target of 1.67 on safety characteristics. The exit criteria matter more than the entry. Containment without an exit strategy becomes culture, and you are still funding it in year five.

Inside that envelope, run QRQC. Every early nonconformance gets an answer within 24 hours, at the line, with the operator who found it – not in a meeting room with people who read about it afterwards. In an established plant QRQC protects a reputation. In a greenfield it does something better: it converts failures into the process memory the plant was born without. A failure at week three, properly dissected, is worth ten assumptions carried over from the design FMEA.

That is also the honest job for AI in year one – capture, not judgement. Use it to log, classify and trace every nonconformance, to build the defect library, so that nothing the process gets wrong is forgotten. Let it earn the right to reject parts only after it has seen a few hundred of them. The defect library, not the robot fleet, is the asset a new plant is actually building.

The plants that age well assumed their own ignorance on day one. Qatar and Daegu will both get their ribbon-cutting photographs, and the machines will look magnificent. The warranty bills arrive three years later, addressed to whoever inherits the plant's memory – or its absence. Novelty gets the ribbon-cutting. Memory pays the warranty.