Benchmarking LLMs on real test-automation work

Twenty models, scored by what their generated tests actually do against a sandbox. They perform exceptionally well at unit-test generation, but stall on the engineering judgment that maintaining a real suite requires.