"AgentShield: Multi-Agent LLM Byzantine Fault Tolerance Evaluation Dataset"

"This dataset supports the paper \"Byzantine Resilience in Multi-Agent LLM Orchestration: Consensus Protocols for Trustworthy Collective Decision-Making\" published in IEEE Transactions on Dependable and Secure Computing (2026). The dataset contains 15,000 multi-agent decision scenarios across four safety-critical domains: medical (5,000 clinical decision scenarios derived from MIMIC-IV under PhysioNet DUA), legal (3,000 case analysis scenarios from CaseHOLD under CC-BY-4.0), financial (4,000 credit and investment scenarios from SEC EDGAR public filings), and autonomous driving (3,000 coordination scenarios from nuScenes under CC-BY-NC-SA-4.0). It includes fault injection configurations for four Byzantine failure strategies (hallucination forcing, prompt injection, stochastic perturbation, strategic adversarial optimization), a curated database of 500 domain-specific Cross-Attestation Probing (CAP) verification probes with expert-verified answers, and sample multi-agent response logs for reproducibility. The dataset enables evaluation of Byzantine fault tolerance mechanisms for multi-agent LLM systems."