LetTokenBurnArrivalBench: Agent-Generated Data Pipelines Are Correct Once and Wrong Under Time
Evaluates agent-generated data pipelines on 40 tasks by re-executing them under adversarial delivery schedules to detect temporal defects like idempotency failures.
$ npm i -g lettokenburn-cli$ ltb bench info arrivalbench
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-5.6 Sol | 7.0 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Claude Opus 5 | 16.7 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Kimi K3 | 35.2 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| DeepSeek V4 Pro | 45.3 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Claude Sonnet 5 | 48.2 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| MiniMax-M3 | 50.2 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Claude Haiku 4.5 | 52.2 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| DeepSeek V4 Flash | 52.4 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Gemma 4 26B | 54.7 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Qwen3 Coder Next | 73.7 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Codestral | 79.2 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
ArrivalBench consists of 40 tasks where agents generate incremental data pipelines. These artifacts are graded not just on a single snapshot but by re-executing them under replayable, adversarial delivery schedules (late, duplicated, out-of-order records). The oracle compares the final state against a batch recomputation of the complete log.
Scoring separates crashes from silently wrong tables. Results show that while 86–100% of pipelines pass standard snapshot tests, 7.0–79.2% of those fail under temporal evaluation, revealing significant idempotency hazards across eleven models.