What stops working
QA as a gate at the end of the line. A manual phase that inspects a machine shaft's output after the fact is a queue in costume; it either becomes the bottleneck or waves machined defects through. The phase dissolves.
What it becomes
The instinct gets promoted, not retired. Quality re-emerges as evaluation engineering: designing the executable checks that let the inner loop verify itself, revolution after revolution. Correct stops being a judgment made at the end and becomes a specification written at the start. Alongside it appears a skill with no pre-AI ancestor — trust calibration: mapping, empirically, where the machines are reliable and where they fail, and tracking how that frontier moves month to month. Over-trust and under-trust are both expensive; only calibration compounds.
The strongest objection
That "evaluation engineering" is just testing with a new name. Contested: testing checked human work after the fact; evals commission machine work before it, and the boundary they police is the only thing standing between confident speed and confident defects.