LetTokenBurnWatts If? A Benchmark for Counterfactual Reasoning About the AI Data Center Buildout
DCBench-440 evaluates LLM counterfactual reasoning on AI data center planning via 200 generative STRIPS questions and 240 multiple-choice questions grounded in site records.
$ npm i -g lettokenburn-cli$ ltb bench info dcbench-440
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-oss 120B | 71.5 | openrouter | Independent | medium | Oct 15, 2026 | arxiv.org |
| GPT-oss 120B | 71.0 | groq | Independent | medium | Oct 15, 2026 | arxiv.org |
| GPT-OSS-Safeguard-20B | 71.0 | openrouter | Independent | medium | Oct 15, 2026 | arxiv.org |
| GPT-OSS-Safeguard-20B | 70.0 | groq | Independent | medium | Oct 15, 2026 | arxiv.org |
| gpt-oss-20b | 68.5 | openrouter | Independent | medium | Oct 15, 2026 | arxiv.org |
| gpt-oss-20b | 65.5 | groq | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3.8 27B | 63.0 | openrouter | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3.8 27B | 53.0 | groq | Independent | medium | Oct 15, 2026 | arxiv.org |
| Mistral-Small-24B | 18.0 | openrouter | Independent | medium | Oct 15, 2026 | arxiv.org |
| Llama-3.1-8B-Instruct | 14.0 | openrouter | Independent | medium | Oct 15, 2026 | arxiv.org |
| Mistral NeMo | 12.5 | openrouter | Independent | medium | Oct 15, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
The benchmark comprises two arms: 200 generative questions over STRIPS worlds for five US deal archetypes with search-verified gold answers, and 240 multiple-choice questions grounded in quoted site records and emission equations.
Scoring uses exact match for generative tasks and accuracy for multiple-choice. Results show open-weight models achieving 12.5–71.5% on generative and 58.3–92.5% on multiple-choice, with universal failure in enumerating available actions.