Apodex 1.1 argues a right answer can still be a failed task. The whole release is built on one claim. The meaningful unit of AI capability is completing a real piece of work end to end, files and code and all, with a result you can check. I went through the release and the technical report, and a few things stood out.