An optimiser found 2,900 cheap designs for a server hall, then 954 of them turned out to fail in the actual building

The task is picking the cheapest fan speed and chilled water temperature for a hall of 20 server racks, and the search sampled 9,000 combinations, of which 2,900 came back feasible. 954 of those 2,900 clear every check against the design plant and then break a constraint on the hall as it was actually built, so roughly a third of the apparently safe answers are fantasy. The sentence in the side panel is part worth keeping, because the gap between plan and reality shows up as a design that fails to run, rather than a design that merely costs more. The run that skipped the site rules came in cheapest on paper at $1,001 a day and then failed as-built, with the inlet temperature landing 0.32 K over allowable, while the run carrying those rules cost $1,008 on paper, $1,021 as-built, and ran. Hand coded penalties, the usual way rules get bolted onto a search, landed at $1,080 on paper and $1,094 as-built, so the version reading a rule book beat the hand-written version on both figures. The transferable part for anyone building a search or an agent is that whatever decides which options count as allowed matters rather more than the optimiser itself, since real system rejecting an answer wipes out whatever that answer saved. Whole run finishes in 15 seconds and costs 40 cents, at 48 designs a generation across 12 generations.