An AI data center’s design locks in capital, water draw, and carbon trajectory years before any watt is actually consumed. Discrete choices—grid vs. on‑site gas, evaporative vs. dry cooling, and the county of filing—determine these commitments. Today, large language models (LLMs) are asked to screen such decisions, yet there is no test of whether they truly capture what each commitment excludes. We introduce DCBench‑440, split into two arms:
- 200 generative questions over STRIPS worlds covering five U.S. deal archetypes, with search‑verified gold answers; no shortcut exceeds 12% on any question type.
- 240 multiple‑choice questions grounded in quoted site records, emission equations, and stated scenarios.
Across open‑weight deployments, generative accuracy ranges from 12.5% to 71.5%, while multiple‑choice accuracy spans 58.3% to 92.5%. One skill fails universally: when asked to list every action available at a decision gate, no model exceeds a 4.0% exact‑match rate. The strongest model lists 72%–85% of actions that would change the project’s state but omits most permitted actions that leave the state unchanged; even counting those omissions as acceptable, the best exact‑match remains 24%.
Code, data, and other materials are available upon request.
Review