Correctness regressions get caught. Performance regressions get shipped, because nothing in a normal pipeline is watching for them.
The setup
Pin the hardware, pin the clocks, run a fixed workload matrix on every commit, and compare against a rolling baseline with a noise band derived from the last N runs rather than a hand-picked threshold.
What it catches
Mostly not the dramatic ones. It catches the 4–8% losses from a driver bump, a framework upgrade, or a kernel change that helped one shape and hurt three others — the kind that compound quietly over a quarter.