There is no posttraining data commons

The open-data ecosystem that enabled language-model pretraining grew from decades of publishing and archiving: Wikipedia, scientific papers, open-source code, and Common Crawl offered vast stores of documents that people had created for other purposes. Posttraining, especially reinforcement learning for AI agents, lacks an equivalent resource. It needs repeatable tasks rather than documents, and each task must preserve a world, a goal, usable tools, and a way to judge success. Real human work is contextual and changes as people act, so converting it into a stable, repeatable task means reconstructing and simplifying activity that may never have been recorded.

Software provides the strongest starting point because repositories, issues, tests, and development environments leave durable traces. Yet even coding benchmarks reveal how much reconstruction is required: only a small share of more than 90,000 pull requests examined for SWE-bench became usable tasks, and most of those required further filtering. In SWE-bench Verified, human reviewers rejected over two-thirds of sampled tasks; other datasets have suffered extensive failures from broken tests, missing dependencies, or leaked solutions. If software is difficult, capturing work in fields such as law, medicine, or science is harder still. The expanding posttraining-data industry exists to manufacture tasks, environments, and graders. Building an open alternative requires systematically turning traces of human work into shared resources - and deciding whether those records become public infrastructure or proprietary assets.