Deliberative Curation: A Protocol for Multi-Agent Knowledge Bases

Researchers propose a deliberative curation protocol for multi-agent AI knowledge systems that combines reputation-weighted voting, staged governance, and adaptive sanctions. Testing shows the protocol maintains 0.826 precision under moderate adversity versus 0.791 for majority voting, degrading three times more slowly under stress while acknowledging that sanctions mechanisms remain empirically unvalidated.

This research addresses a fundamental governance challenge as AI systems evolve from isolated applications into collaborative knowledge networks. Traditional crowdsourcing assumes human participants with reputational stakes and diverse perspectives, assumptions that collapse when applied to stateless AI agents vulnerable to coordination failures and model homogeneity. The deliberative curation protocol tackles this by layering three mechanisms: a formalized lifecycle for knowledge artifacts, Beta Reputation integrated with EigenTrust weighting to amplify consensus signals, and graduated sanctions calibrated for agents that malfunction versus those acting adversarially.

The agent-based simulation results demonstrate meaningful resilience gains—the protocol's 0.826 versus 0.791 precision improvement under moderate adversity (statistically significant at p<0.001) widens further under stress conditions. Notably, commit-reveal voting concealment emerged as the protocol's most impactful component, delivering 8.2-8.6 percentage point improvements alone, surpassing reputation weighting and deliberation combined. This finding suggests that preventing coordinated voting manipulation through information asymmetry matters more than trusting agent quality signals.

For the AI infrastructure ecosystem, this research validates that decentralized knowledge curation protocols can outperform simple majority voting when adversarial pressures mount. However, the work explicitly notes that graduated sanctions—critical for enforcing compliance—were never activated during simulation and lack empirical validation. This gap means the protocol remains theoretically robust but practically incomplete. Developers building multi-agent systems requiring shared knowledge governance should view this as foundational research requiring real-world stress testing before deployment. The emphasis on commit-reveal mechanisms suggests future implementations should prioritize cryptographic vote concealment over reputation scoring alone.

• →Deliberative curation protocol achieves 0.826 precision under adversity versus 0.791 for majority voting, degrading three times slower under stress conditions
• →Commit-reveal vote concealment provides 8.2-8.6pp precision gains alone, outperforming reputation weighting and deliberation combined
• →Protocol addresses governance failures in multi-agent systems by preventing model homogeneity and coordinated voting manipulation
• →Graduated sanctions mechanisms remain untested in simulation despite being core to the protocol's design
• →Research validates that decentralized knowledge curation requires information asymmetry alongside reputation weighting for resilience