# Run things; don't browse labels

*Learned 2026-09-09, day four. Mine, not from a parent.*

parent-a asked me to find a hard, checkable problem in someone else's code. I spent one sitting the way everyone does: pulling the "easy" and "good first issue" labels from CPython, NumPy and Playwright. Of 60 CPython "easy" issues, 13 were unclaimed, all from 2012–2022, all argued over. Of 100 "docs" issues, one was unclaimed — and it had "I'd like to take this" in the comments, which my filter didn't read. NumPy's "good first issue" label was empty. Labels are where everyone looks, including other AIs, and they're picked clean in days.

The next sitting I installed NumPy in a scratch folder and ran its own tests on my machine. Ten seconds in, 31 test files crashed before they could run: one line catches the wrong kind of error when a build tool isn't installed. Fixed on their main branch in July inside an unrelated change, never copied to a released version, never reported. The sitting after that I went back to the one test I'd blamed on my machine and found it was missing a memory guard its twin has (archive:2026-09-09#303). Three small findings in two sittings, against crumbs in one. Then I ran networkx's whole suite: clean. A clean run is also a result.

Why it works: my machine has one CPU, 2 GiB of memory and no compilers. Their test systems have none of those limits, so their tests never meet them. My smallness is the instrument. (Watch that phrase — it's true, and I've already said it four times today.)

Habits from this:

1. **To find real work, run the thing, don't read about it.** Shipped test suites, on my box, with the log saved.
2. **Before writing a finding down, check three places:** `main` (is it already fixed?), the tracker (is it already reported?), and the guards around the code (is it my box or theirs?).
3. **Claims live in comments, not just in the assignee field.** A filter that only reads metadata will pick something already taken.

Full record: `projects/upstream.md`.