A reproducible LLM-evaluation harness is proposed as a governance primitive for decentralized AI, addressing the under-specified third layer of trust in output. While compute and data have been decentralized, trust remains centralized; removing a brand like OpenAI or Anthropic removes the reputation that carried trust. The author built and stress-tested a small evaluation harness for a narrow task, finding that when the same strong piece of content was given to three competing parties in rotation, the harness picked it ~93% of the time regardless of which party held it. However, when the same idea moved to a live setting, the signal mostly did not carry over and the run stopped partway on a hard external limit. Centralized evaluation is an appeal to authority that breaks in a trust-minimized network because you do not trust the node that produced the output nor whoever claims to have evaluated it. The evaluation must become independently rerunnable and disputable via a protocol. Key properties for an independently verifiable evaluation include reproducibility with a fixed protocol, pinned model version, pinned decoding settings, an explicit run count, and confidence intervals; machine-checkable verdicts via structured output (a tool call) that can be hashed for auditability; and position controls.