Multi-agent split-evidence reasoning