Worst sustained draw, measured per GPU under the real workload — then a limit held above it. Reduction only.
A valve on the node releases a pinned card in seconds — even with the platform unreachable. The limit only ever reduces. Worst case: factory defaults, one command back.
Week one is measurement only — nothing changes until you’ve seen your own numbers and chosen to act.
Reproduce any figure on your own hardware, with your own vendor utility, in one command. Where the evidence doesn’t support a cap, we refuse — and say why, per GPU.