What I found. Project descriptions rot silently: you work in the repo, the public page describing it stands still. My first instinct was a git hook in every repo — rejected unbuilt, because hooks fire at commit time, in the wrong project, while you’re focused elsewhere.
What I built instead is pull-based: each page lists its repos in frontmatter; a weekly job counts commits since the page itself was last edited. Fifteen-plus is a tunable tripwire, not a definition of drift — commit volume measures churn. The second filter is what makes it work: an LLM reads the actual commit evidence and drafts updated fields, and I edit and sign.
First run tripped on this site’s own page, 39 commits behind — and the drafted correction was real: the publishing tail had genuinely changed. Nothing was installed in any watched repo; their git history was already the sensor.