Once AI sits inside a machine’s safety architecture, compliance stops being the last stage and becomes a design constraint. René walks through what the Machinery Regulation already demands of self-learning safety functions, bounded task envelopes, recorded safety decisions, third-party assessment, and the four items that belong in the first architecture review: task envelope, non-bypassable log, controlled update path, examinable override. Plus the integrator trap: fine-tune or modify, and you may become the manufacturer. Full article at renegrywnow.com.
Reflection Questions
Who in your organization decides what the physical AI is allowed to learn, log and do?
If you fine-tune a model next quarter, do you know whether you just became the manufacturer?
Is your safety case an architectural property — or a document someone writes at the end?
