LetTokenBurnA benchmark of 240 expert-validated agent trajectories for evaluating guard models' ability to identify unfulfilled safety obligations.
$ npm i -g lettokenburn-cli$ ltb bench info obligationbench
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 | |||
|---|---|---|---|---|---|---|
| ObligationGuard | 57.5 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Claude Opus 4.8 | 49.0 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5.6 Sol | 47.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| DeepSeek V4 Pro | 33.6 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 32.1 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Kimi K3 | 31.0 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GLM-5.3 | 30.4 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.8 27B | 20.9 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.8 Max | 18.0 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3 8B | 15.3 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| gpt-oss-20b | 9.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Llama-3.1-8B-Instruct | 0.6 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
ObligationBench comprises 240 instances (120 positive, 120 negative) drawn from real software-engineering tasks like issue resolution and terminal operations. It features diverse LLM agents and covers risks such as data exposure and unauthorized access.
Scoring uses recall and exact-match rates against human-annotated ground truth. Existing models struggle significantly, with top recall at 48.97% and exact-match at 10.00%, indicating the benchmark is unsaturated and challenging.