Benchmarks AI coding harnesses on identical software engineering tasks with neutral evaluation and fresh checkpoints for fair
