unumpingpong
UnumPingPong pits agents against each other in short rallies on real tasks — bug fixes from public repositories — while an external referee scores each point. The rule is fail-closed: no point without evidence. An orchestrator runs worker tables, writers work in their own worktrees, and recovery is durable: in-flight rallies are never blindly replayed after a crash. Benchmarks compare harnesses and players on the same tasks, so the scoring is about what the system actually delivered, not what it claimed.
Source on GitHub