Contamination Percolation in Multi-Agent LLM Systems: A Measurement Framework and Benchmark

Multi-agent large language model (LLM) systems are being deployed in healthcare, finance, and legal settings, yet reliable methods for measuring misinformation spread through them are lacking. We introduce ContamPerc, an initial benchmark of 400 vignettes across 10 domains (120 author-validated, 280 community-released), 50 domain-specific semantic markers, and a measurement pipeline that separates genuine contamination from defensive citations. The benchmark defines the Contamination Gap Diagnostic (CGD), a single number that, read with its two component rates, classifies a model’s alignment behavior; it is not monotonic in safety: near-zero can mean blocking both payload types or propagating both. Rigor rests on the design, not the scale: 100 trials per cell with Wilson 95% intervals, a bootstrap test ( B = 10,000 ), isolation and within-family size controls, a human-labeled detector check, and a real-case validation slice, across five model families (GPT-OSS 120B, Claude Sonnet 4.6, Llama 4 Maverick, Gemini 2.5 Flash, GPT-4o-mini) and roughly 246,000 API calls. The CGD reveals five distinct contamination profiles spanning +51 to −64. The spectrum alone understates the risk: plausible clinical misinformation propagated at up to 92% under full connectivity, and even the model that rejected nearly all nonsense control tokens propagated it at 60%. Sensitivity analysis across three temperatures and two prompt phrasings confirms robustness; cross-domain validation on legal and financial vignettes across all five models is consistent with generalization beyond healthcare. On a 100-case real-clinical MIMIC-IV slice run on all five models, four retain their synthetic categorical classification under full connectivity; the near-zero model (Claude) shifts to instruction-following, confirming a boundary-sensitivity prediction. Tier-1 magnitudes are 15–25 points lower on real cases for four of five models, consistent with synthetic substrate providing upper-bound contamination estimates.