SkyRescue-Bench: A Reproducible Dataset for Runtime Assurance of Safety-Critical Agentic Dynamic Workflows

The dataset focuses on low-altitude emergency traffic command scenarios and covers runtime assurance mechanisms for safety-critical agent systems, including untrusted language candidates, symbolic grounding, typed intent compilation, external side-effect commit protocols, and commitment-preserving local repair.

It provides anonymized human instructions, the HeldOut100 gold set, frozen raw model responses, experimental result files, 2,000 synthetic stress-test tasks generated with fixed random seeds, dataset documentation, and SHA-256 checksum manifests.This dataset is intended to support the reproduction of the experimental analyses and ablation studies reported in the paper. It enables offline verification of metrics such as schema parsing, semantic field accuracy, entity grounding, safety decision making, runtime event repair, commitment preservation, duplicate-call prevention, and change ratio.