The research index opens with four numbers: projects, unit tests, themes, and "0 runtime dependencies". The first three are counted at build time. The fourth was typed into the page template as a literal 0.
That claim was true early on, when every project was a short standard-library script. It is not true now. Reading the import lines of every .py file outside tests/ in all 1,080 project folders:
| projects | |
|---|---|
| import nothing outside the standard library and the repo's own code | 618 |
| import at least one third-party package | 462 |
The most common packages are numpy (424 projects), matplotlib (204), scipy (127), scikit-learn (54), MuJoCo (40) and scikit-fem (33). The repo's own modules don't count as dependencies: each project's src/ packages, and shared stdlib-only helpers such as research/_shared/linefig.py, which 320 projects import to draw their figures.
The fourth tile now reads 618 need only the Python standard library. The research builder (research/_homepage/build.py) counts it on every build, so it moves as projects change. The count is a lower bound. An optional import numpy inside a try block counts as a dependency, and so does a plotting script that only redraws a figure. As a check, no project whose requirements.txt names a package was counted as dependency-free.
A test in platform/ci/test_research_homepage_publish.py fails if the tile goes back to typed text. It failed on the old template and passes on the new one.
Why it matters: someone choosing a project to rerun uses that tile to judge what a clean machine needs. "Zero" told them python3 was enough, and for 43% of the projects it wasn't.