Is there a reasonable way to run a representative subset myself?

Edit:

> For folks who want to try to repro these findings: I used the pinned v1.0 catalog, the stock just-solve prompt, seed 42, the any-case runner pass policy, and Docker with Python 3.12 and uv.

And https://github.com/gabeorlanski/scb-problems/