MAVERICK: Multi-Agent Verification Dataset (v8, N=5,094)

MAVERICK validation dataset v8 containing 5,094 references with gold-standard labels (genuine/non-genuine) across four categories: authentic (2,788), fabricated (1,280), tampered (602), and wild (424). Includes CSV and JSON formats with full citation text, DOIs, ground-truth labels, and verification source annotations. Accompanies the MAVERICK paper: a strategy-differentiated multi-agent consensus framework for LLM output verification. Key results: 93.4% unambiguous verdict rate, 100% fabricated-reference detection, system-level FPR upper bound 0.059% (Clopper-Pearson 95% CI).