Air traffic control for AI coding

The coordination benchmark, adjudicated.

A vendor-run benchmark of Nool against git-based coordination baselines, built so a skeptic can check every number: each claim is stated so it could fail, with the evidence, and with what it does not establish.

• Run 9 (nool 6.14.1, 2026-08-21, synthetic agents, 10 seeds): bare Nool 718.5 vs. 635.4 valid tasks/hour against git worktrees, 77 vs. 444 child processes spawned.
• Full governance ties git worktrees on throughput (647.7 vs. 635.4, not significant after Holm correction) but adds about 0.2 s of average stall.
• Fast mode does not govern: none of the three built-in recovery checks catch a change that already landed. Measured on 6.11/6.14; a 7.10.0 re-run is scheduled.
• Mechanism results on 7.8.0–7.9.2: zero false positives in blast radius across 30 languages, admission cost growing 3.5–9.7× over a 100× larger repository, and per-operation latency 11–79× git's.