Can Frontier AI Build a Statistical Register? A Benchmark and Research Programme on Vietnam's Thermal Power Fleet

Energy policy models need asset-level data that are complete, current, historical and prospective, and traceable to authoritative primary sources. Yet across much of the world these data arrive late or incomplete. We introduce a computable benchmark for this task: reconstruct the register of Vietnam's 177 thermal power plants and score the result against a hand-compiled, per-cell-sourced reference. We find that the task is hard for today's AI systems. Fourteen language models answering from memory recovered much less than the Wikipedia lists contain, even though these lists had presumably been in their training corpora for years. We observed that a reference-free signal, the within-run variability of the reported capacities, correlates with accuracy, letting us screen out weak runs without the reference. We then tested the commercial frontier of mid-2026: four agentic systems with web access and extended reasoning. Only one improved coverage over the memory-only query, and a multi-step harness degraded three of the four. A curated document set improved coverage, most for the weaker agents. Fusing the lists from multiple runs more than doubles recall. We conclude that an evergreen statistical-quality register requires deliberate knowledge engineering. We propose deriving datasets as dated snapshots from a sourced, auditable knowledge base, mechanically updated from a periodically harvested corpus, with humans in the loop to vet sources and resolve the hard tail.