Measured

An AI data center is a stack of optimization problems: where the racks go, how equipment is placed, the GPU kernels, and the inference bill. Code can check every one. Our harness turns any problem with a programmatic verifier into a team of Nemotron agents on Crusoe that improves itself. Agents share an evidence-gated memory: a lesson enters only if it cites a measured result. After each round, a research agent proposes changes to the team as papers, each with a predicted gain. The verifier reruns the dev tasks, and a change is kept only if it measurably scores higher or cuts cost by 15% or more. No LLM grades anything. The final version is scored on held-out tasks the team never saw. We ran it on four layers. Inference cost per correct answer fell 31%, with no answer lost. Rack placement improved from 0.597 to 0.845. The gap to QAPLIB's best-known solutions narrowed from 4.3% to 3.6%. Correct GPU kernels went from 1 to 2 of 5. On hall layout, the gate rejected every proposed change, because none measurably helped.