Notes
Working notes on method.
Occasional writing about how expert data actually gets made — reward design, verification, and the awkward economics of persuading busy professionals to do careful work. No announcements.
22 Jul 2026
Write the grader first
If you can't say how you'd check the answer, you don't have a task yet — you have a wish. On why reward design has to precede data collection.
Read → 19 Aug 2026The expert supply problem
The people you most want writing training data are the people least willing to do it. What actually moves them, and what doesn't.
Read →Disagree with something here?
We'd rather hear it than not. Arguments about method are how the specs get better.