TL;DR — Controlled empirical measurement of adversarial verification for multi-agent fan-out: the cheapest tier (Haiku, N=1, $3.46) Pareto-dominates all 27 configurations, and increasing skeptic breadth beyond N=1 can harm safety due to a Condorcet-style vote-dilution effect.
ThakiCloud AI Research · 2026-07-11 · 📝 Tech blog (KO)… See the full description on the dataset page:
https://huggingface.co/datasets/thaki-AI/daily-paper-2026-07-11-verify-gated-fanout-pareto.