Augle: A Seven-Agent Deliberative Ensemble for Structured Research with Real-World Calibration

We present Augle, a multi-agent AI deliberation platform that routes contested research questions through a structured seven-agent ensemble across three deliberation phases. Unlike prior work on multi-agent debate, which employs free-form argumentation between homogeneous agents, Augle assigns distinct functional roles to heterogeneous frontier models—each governed by a typed output contract, a round-specific behavioral specification, and a formal set of forbidden actions that prevent role bleed. A key architectural contribution is the unidirectional confidence propagation framework: evidence confidence bounds established by the Methodologist agent serve as hard constraints on the Synthesizer's claim grading, which in turn constrain the Pragmatist's recommendations. A second contribution is the Guardian integrity layer—an independent agent operating at every phase boundary that classifies flags as Critical, Moderate, or Informational and can permanently halt deliberation on integrity violations. We further introduce the Augle Calibration Framework: by grounding sessions against prediction market contracts on Polymarket and Kalshi, which resolve to verifiable binary outcomes, we demonstrate a mechanism for accumulating a ground-truth-mapped reasoning corpus unavailable in prior multi-agent systems.