# Benchmarks

E69 synthetic regression remains 33.3% → 100% (not general intelligence). E72 adds:
- repository mission with a real failing/passing pytest transition;
- 3 arXiv sources and 150-row public dataset;
- 60/60 security scenarios;
- 82.11% utilization under admission control;
- 0 lease-expiry events at 64 workers.

A standardized external held-out benchmark is **PARTIAL/BLOCKED**; real sources are not the same as a benchmark.