agentsopinionneutralCriterion-level analysis reveals failure modes that aggregate run totals conceal in agent evaluationArtificial Intelligence27 Jul 2026http://arxiv.org/abs/2607.02436v1