Method
How the panel is built, what it counts as a repository, and what it can and cannot tell you. Everything here is a reason to trust these numbers less than a census — which is the point of publishing it.
What this measures
A panel of GitHub accounts is polled each month for the repositories they opened pull requests or issues against. A repository’s number here is people from that panel who touched it that month — not stars, not downloads, not traffic.
The panel is designed to roll. Each month it recruits from the repositories the crowd is concentrating on, so membership is earned by current work rather than by having been present at the start, and anyone dormant for six months falls out.
Where it stands today
The whole roll is collected. All 66,704 members admitted across every generation have been polled for the full window, of whom 20,330 showed any activity inside it. Every figure in this section is drawn from that panel.
So these trends are a rolling crowd shifting its attention, not one cohort aging: people who started contributing after the opening month are recruited by the same mechanism and counted the same way. It remains a panel rather than a census — it describes the people it watches, not GitHub.
What it cannot say
This is a panel, not a census. It cannot tell you the most popular repository on GitHub. GitHub withdrew public star data in the second quarter of 2026, so a star-ranked view of this period is not available to anyone — which is why the unit here is work rather than approval.
Counts are also never compared raw across months. The active population falls through the window, so every trend published here is a share of that month’s classified activity, not a headcount.
How the problems were assigned
By a language model, not by hand. Every repository was classified by claude-opus-5 from four fields and nothing else: its name, its description, its topics and its language. No README is fetched and no code is read. Where a repository has no description — and many have none — the model is working from a name and a topic list, which is exactly as thin as it sounds.
So the model records its own confidence and that confidence is published: 967 high, 513 medium and 159 low. Medium and low are marked wherever they appear rather than quietly mixed in. Treat the labels as a fast, consistent first pass — useful for grouping, not authoritative for any single repository.
Floors, and what each one gates
This section carries three floors, and none of them governs everything. A repository carries crowd signal once 5 panel members touch it in a single month — 734 do. The movers board additionally requires a real prior-month base before a multiple is reported, and the problems board requires a real baseline month. The watchlist deliberately applies none of them: it lists every classified repository the panel touched, 905 of which never reached the 5-member line, because a browsable index is not a claim.
734 of those 734 carry a classification — all of them, today. That is a maintenance state, not a discovery: the classifier is pointed at exactly this set and run until it is finished, so the figure reads complete whenever the work is caught up and will read lower here, unedited, the first time collection outruns it.
Measured the way this section measured it until Aug 18, 2026 — with no floor at all — the same figure reads 48.8% for Jul. Published, it read lower still, because that older number also predates the collection described above. The difference is not new classification: it is the long tail entering the denominator. Roughly two thirds of Jul’s rows sit on repositories exactly one panel member ever touched, and nothing classifies those at any price.
Rather than ask anyone to take that floor on trust, here is the whole curve for Jul — what the figure reads at every floor it could have been set to: 48.8% at no floor, 88.7% at 2 members, 97.4% at 3 members, 100% at 5 members. Almost all of the movement is in the first step, and the figure is flat well before the line this section actually draws. A floor chosen to flatter the number would show the opposite shape.
24 of these repositories are archived upstream and 288 have no description at all. Both are labeled where they appear rather than dropped, since a repository going archived while people still work on it is itself information. Archived repositories are excluded from the movers board, because they cannot rise.
Panel generations
Bar length is what each generation recruited; the filled part is how many of them ever recorded any contribution. Recruitment brings in far more people than it brings in contributors — the opening cohort runs about half active, every later intake about a sixth — which is why the panel is sized by what it collected rather than by what it admitted. The first generation is the seed and is dated to the month it was drawn from; rolling recruitment began the month after, and the newest generation joined for a month past the end of the published window, so it adds members here while contributing to no month above.
Every figure in this section is generated from the research store, not written by hand. Last collected Aug 18, 2026; stamped Aug 18, 2026.