Harness evolution reportedly improves game-benchmark results but doesn't beat test-time scaling on Terminal-Bench

In an October post, the author says harness evolution does not beat test-time scaling on Terminal-Bench but reports +80% on ARC-AGI-3 and +11% on EdgeBench games under matched budgets. The post suggests it can help adaptation in long-horizon, out-of-distribution games. In July, the same author said its gains largely came from repeated sampling and showed limited generalization.