Title: Deliberative Curation: A Protocol for Multi-Agent Knowledge Bases

Abstract

As artificial intelligence agents evolve from standalone utilities into collaborative entities within shared knowledge ecosystems, the management of collective knowledge curation has emerged as a pivotal challenge. Traditional human-driven platform governance models are ill-suited for this new context because agent statelessness neutralizes deterrence-based penalties, model uniformity breaches the independence assumptions required for crowd wisdom, and sycophantic tendencies erode deliberative consensus. To address these issues, we introduce a deliberative curation protocol structured around three distinct governance layers. First, we formalize the lifecycle of knowledge artifacts using a labeled transition system. Second, we implement reputation-weighted deliberative voting, which merges Beta Reputation with EigenTrust amplification. Third, we design graduated sanctions tailored for stateless agents, featuring a mechanism to distinguish between agent malfunction and adversarial conduct.

We assessed the protocol’s efficacy through agent-based simulations involving 100 agents representing seven behavioral archetypes, tested against two adversity scenarios with 30 seeds each, utilizing paired t-tests. The results indicate that while the protocol exhibits slightly lower precision under benign conditions, it offers significantly enhanced resilience during adversity. Specifically, under moderate adversity, the protocol achieved a precision of 0.826 compared to 0.791 for majority vote (p<0.001). This performance gap widened under stress, with the protocol reaching 0.807 versus 0.740 for majority vote (p<0.001). Furthermore, the protocol’s performance degraded approximately three times more slowly than that of majority voting. An ablation study revealed that commit-reveal vote concealment was the most influential single component, yielding an 8.2-8.6 percentage point improvement in precision (p<0.001), thereby outperforming the combined effects of reputation weighting and deliberation. It is important to note that graduated sanctions were not triggered during the simulation and thus remain empirically unvalidated.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC