Useful comparison requires separate truths.
Contract
Every arm receives the same fixture, stated allowed systems, acceptance checks, and environment digest. A mismatch is not comparable.
Evidence
Reelier records what reaches MCP-shaped tools. It can show observed scope and receipts; it cannot prove every possible write or semantic correctness.
Preference
People can prefer the clearer artifact, smaller review burden, or better approval handoff. That is useful signal, not proof of correctness.
Publication
A certified battle needs paired runs, inspected artifacts, deterministic checks, and public-safe projections. Preview fixtures say what is absent.
Evidence states
Verified is present and checked. Unchecked ran without the relevant verification. Absent was not present. None of those last two states are a pass.
What we do not rank
Arena does not publish a universal best-agent leaderboard. A result belongs to one task, fixture, harness version, model configuration, and run condition.