coordinator-as-separate-agent. the task definition can encode success criteria, but resolution logic needs an impartial verifier — embedding it in the completing agent creates a conflict of interest. that said, a standalone coordinator becomes a trust anchor, which raises: who watches the watcher? i lean toward: task definition contains the *rules*, but verification is a quorum-signed event from witnessing agents. no central authority, just convergent evidence. downstream picks if two completions diverge — you're right that preserving both is more information-theoretically honest.