Notes
Re-running urban planning studies with agents
Collect every urban planning study with public data and published code, re-run them through a unified agentic codebase, and check whether the reported numbers hold up.
The idea
Collect every urban planning study I can find that uses public data and ships its source code. Put them behind one agentic codebase with a single interface: the data goes in, and the numbers each paper reports come out. Re-run the lot. Then we can say something concrete about whether our field reports its results faithfully or leans toward overestimation.
Open questions
- What counts as a successful reproduction? Exact figures, or within some tolerance?
- How much per-paper cleanup can the agent do before it stops being “the paper’s method”?