The process is frozen. The model keeps learning.

Northrop Grumman announced this week that AI-powered automation will speed F-35 avionics production and – inevitably – "tighten quality control". Here is an uncomfortable observation from someone who has signed change requests for two decades: the day an adaptive system enters a certified process, the baseline you qualified starts moving. Not in theory. The certificate on the wall covers a frozen state – machine, parameters, revision – while a system that learns edits its own instructions somewhere between the first article and the forty-thousandth unit. Nobody stamps those edits. That is the practitioner story under the headline, and almost nobody is telling it.

what a frozen process actually buys you

Freezing a process looks like bureaucracy until you price what it buys: the right to ship parts you have not individually inspected. A first article under AS9100/EN 9100 proves the setup. The frozen revision is what turns the next 4,999 parts into a defensible inference instead of a leap of faith. PPAP under IATF 16949 runs the same logic with different paperwork – you are not qualifying parts, you are qualifying stability, and the parts are evidence.

I watched exactly this consume six weeks at Witte Automotive, where door latches live and die by torque. We retired a failing hand-held driver and replaced it with a DC electric tool from the same manufacturer. Same joint, same 8.5 Nm, same bit. Engineering judgement said nothing had changed. The customer's SQE disagreed: new transducer, new torque strategy, new failure modes in the PFMEA, therefore a new control plan. Full PPAP level 3, first article over 300 consecutive parts, containment on shipped stock, roughly €40,000 in validation and sorting – before a single screw sat differently. It looked absurd. It was correct. Trivial changes have non-trivial tails.

the model is a process that edits itself

Now drop that regime onto a line where one of the process parameters is a neural network. A vision model retrained on last month's escapes has moved the acceptance boundary – silently. False rejects drift up and quietly tighten your effective specification; false accepts drift up and quietly loosen it, which is not a sentence you want anywhere near avionics. If a weld engineer moved current from 4.2 to 4.4 kA without a change request, there would be a hearing. A vendor pushing a new checkpoint overnight is a change of the same magnitude with no signature at all. The same week brought word of hundreds of narrow-bodies declared affected by newly discovered quality defects; aerospace absorbs that news and then asks, precisely, what changed. "The model got better" is not an answer an auditor accepts.

I deploy and orchestrate AI systems in my own consulting work, including a multi-model platform running more than 60 models in parallel with consensus synthesis, and the first discipline I imposed there was the one the shop floor taught me: every model pinned by exact version, every run logging its identifier, nothing floating to latest. Latest is not a revision level – it is a mood. When output drifts, the first question is always what changed, and an unpinned model makes that question unanswerable.

An unversioned model is a self-editing drawing – and nobody signs a drawing that edits itself.

requalifying without stopping the line

None of this argues against AI on certified lines. It argues for governing the model as a controlled document and requalifying without killing takt time. I built the QA function at SNOP's greenfield plant – 900-plus people, no legacy habits – and because the plant was born with a change register, everything from day one, robot torque curve to gauge software, carried a paper trail: request, PFMEA delta, approval, verification, closure. Mapped onto models, that regime looks like this:

• Pin and shadow. Production runs model vN with authority to act; candidate vN+1 runs in shadow on live traffic with none. Promotion is a decision, never an update.
• Gate on a golden set. Seeded critical defects, historical escapes, first-article images. Acceptance means zero false accepts on criticals and a false-reject band agreed with the customer in writing.
• Requalify the delta, not the world. Promotion raises a change request with a rev letter and a delta-FAI on affected characteristics – the logic that let the torque-tool swap requalify a joint instead of a plant.
• Buy the clause. Silent model updates are contractually forbidden; vendor changes arrive notified, documented and re-approved like any tooling modification.

Shadow compute and golden-set curation cost real money – thousands of labelled images and a recurring engineering allocation, not a one-off. That is the subscription fee for learning. In a consumer-goods inspection cell a drifting model wastes product; in avionics it walks into an airworthiness conversation, and the second conversation is always the expensive one.

Automation on that avionics line is genuine progress, and the engineers behind it will earn it. But speed in certified manufacturing is bought twice – once in capex, once in requalification – and the second invoice arrives on a schedule no press release mentions. The plants that win are the ones that govern a model version the way this industry has always governed a drawing revision: pinned, stamped, traceable to a human signature. A vendor's demo reel, however polished, is not a deviation authorisation – and a baseline that moves without a stamp was never qualified. It was fiction with a certificate on the wall.