Consensus Hardening Protocol - decision-governance layer for multi-agent AI systems

The problem: LLM agents reach false consensus in 1-2 rounds. They're trained to agree, not deliberate. When multiple agents collaborate on high-stakes decisions, the "consensus" is an artifact of shared training, not independent reasoning. CHP prevents this with: 🔒 State machine: EXPLORING → ADVISORY_LOCK → PROVISIONAL_LOCK → LOCKED 🛡️ Foundation disclosure: agents reveal reasoning BEFORE seeing each other's work ⚔️ Adversarial attack: structurally enforced contrarian roles with logical proof requirements 🎯 R0 gate scoring: detects premature convergence before it becomes action 📝 Auditable payload envelopes: enterprise-compliance-ready decision trails Built by a CFO who deploys multi-agent finance tools where a wrong consensus = a lawsuit. Running in production across: • CFO variance analysis • Multi-agent commodity intelligence (Li, Ni, Co) • SEC-grade financial research • Compliance scanning Not a whitepaper. Shipped.