Before teaching an agent to compare suppliers, check whether each quote says if freight is included. WorkFactory maps requested results back to the information they need, then has a separate probe attempt the task using only the available files. Repairs supply missing premises without giving away the answer.

The October 27 preprint reports 112,686 tasks across 612 occupations and improvements on four professional-work benchmarks. The evaluation is text-only, and its WorkPilot Elo ratings are local estimates, not official leaderboard placements.

A practical release caveat: the public repository currently contains documentation and assets, with placeholder dataset/model links. Read it as a research method, not a ready-to-run training kit.