The Open-Source Momentum Screen: Category Repos to a Weekly Founder Queue
Run a weekly screen that turns public GitHub events in one category into 5 to 15 de-noised, unfunded founder leads with dated triggers, in under three hours.
Key takeaways
- Peer-reviewed research on 160,065 US startups found that GitHub activity speeds up in the twelve months before a first round and drops in the twelve months after, so the outreach window is months long, not weeks.
- Measure the change in rate over the latest 90 days against the 90 days before, and queue a repo only if stars, distinct outside human contributors and commits all rise together.
- Fake stars stop promoting a repo in under two months, so any month where more than 10% of new stars match the StarScout low-activity or lockstep pattern gets the repo dropped, unless it was a launch burst and the repo kept growing past 60 days.
- GitHub has closed the stargazer list to non-admins, so GH Archive WatchEvents are now the only public source of stargazer identities for fake-star screening.
- In Refolk's index, only 4 US founder profiles mention "open source LLM" in the headline, and 16 to 24% of title-filtered samples are not founders, so finding founders has to start from the repo and work outwards.
- More than half of private raises never get a Form D filed, so a funded-status check also needs company announcements, founder profiles and the CRM.
This is the weekly pipeline for finding open-source teams before their round is public. You pull every repo in your thesis categories from public event data, score sustained growth across several metrics at once, strip out bought stars and one-off spikes, link the survivors to a company and named founders, and drop anything already funded or already in your CRM. It is for early-stage investors, platform and talent partners at funds, and angels who want to find open-source startups on GitHub themselves rather than buy a feed that ranks on one metric. Run as written, it produces 5 to 15 dated, cited queue entries in under three hours a week.
What the weekly screen produces and why the timing works
The screen produces a weekly queue of 5 to 15 open-source teams. Each entry has a named legal entity, one to three named founders, a dated trigger event and source URLs. The timing works because, in peer-reviewed data, startups' GitHub activity rises in the year before a first round and falls after it.
The evidence comes from Roche et al. in Organization Science, which linked public Crunchbase records to the GitHub accounts of 160,065 US startups. Four findings shape this screen:
- Startups speed up their GitHub activity in the twelve months before their first round.
- Among funded startups that are active on GitHub, 80% engage with open-source communities before they are funded, and on average their GitHub activity starts five months before the first round.
- Startups active on GitHub are 15 percentage points more likely to have raised a round.
- After a round, their GitHub activity drops over the following twelve months.
The practical point: the window for reaching out is months long, not weeks. One vendor has published its own claim, based on 4,200 startup GitHub orgs, that a joint jump in commits, contributors and new repos comes before Series A announcements about 70% of the time, with a 3 to 6 week lead. The method is not published and has not been peer-reviewed, so I treat it as unverified. The useful inverse is solid, though: a repo whose activity is slowing after a burst has probably already raised.
This screen finds candidates. Judging one repo's venture potential in depth is a separate read, done after a team reaches the queue.
The weekly momentum screen
- PullSeven daily GH Archive tables for every repo in the category
- Score90-day change in stars, outside contributors, and commits
- De-noiseDrop months above 10% fake-star share and one-block spikes
- ResolveRepo to legal entity and 1 to 3 named founders
- ClearFunded-status and CRM check on every candidate
- Queue5 to 15 dated, cited entries
The weekly procedure, step by step
The procedure has seven steps. After a one-time 30-minute setup, it takes about two and a half hours a week. Each step has a fixed output, so you can hand it to an analyst and check the work.
Category repos to a weekly founder queue
- Define the category universeSet topic, keyword, and org seed lists and fix them as a list of repo IDs with written inclusion rules. About 30 minutes once, then 5 minutes a week to add new matches.
- Pull seven days of eventsDry-run, then run a GH Archive query over the last 7 daily tables for WatchEvent, PushEvent, PullRequestEvent, and ForkEvent, selecting only needed columns, under a byte cap. About 15 minutes.
- Score 90-day accelerationCompare the latest 90 days with the prior 90 on stars, distinct outside human contributors, and commits, and keep only repos where all three rise. About 20 minutes.
- Strip fake stars and hype spikesApply the StarScout low-activity and lockstep signatures to WatchEvent actors and drop any repo with a month above 10% flagged, routing launch bursts to review. About 20 minutes.
- Resolve each survivor to a company and foundersCheck org domain, commit author email domains, license holder, and registry owner, then search for people. About 45 minutes, ending with a legal entity and 1 to 3 named founders each.
- Check funded status on every candidateSearch Form D on EDGAR, company announcements, founder profiles, and your CRM, and log a drop reason for every removal. About 20 minutes.
- Write the dated queueWrite 5 to 15 entries with a dated trigger, the metric changes, the worst-month flagged share, and source URLs. About 20 minutes.
The time budget adds up to 145 minutes a week: 5 + 15 + 20 + 20 + 45 + 20 + 20. Finding founders takes the longest, and it is the step with no published standard method, so if you run over, this is where it will happen.
For step one, write the inclusion rules down as you would a search query: the topics, the keywords in repo descriptions, and the orgs you seed by hand. Save the result as repo IDs, not names, because repos get renamed and moved between orgs. I set no star minimum for joining the universe. ROSS only starts measuring at 1,000 stars, and this screen exists to find teams before that point.
Where the star, commit, and contributor data comes from
Use GH Archive in BigQuery as the single source for events. It is the only public route that still gives you the identity of every person who starred a repo, and the fake-star filter needs those identities. GitHub has limited its stargazer list endpoint to admins and collaborators. The newer history endpoint returns counts without identities.
| Path | Paging | Who can read it | Gives stargazer identities |
|---|---|---|---|
| Stargazer list endpoint | 100 per page, stops at page 400 (40,000 stars) | Admins and collaborators only | Yes |
| Stargazer history endpoint | 30 weeks per page, 100 pages | Any public repo | No, counts only |
| GH Archive WatchEvent | No paging, billed per byte | Public | Yes, actor ID |
The restriction also reaches the web view. Third-party star-charting tools now return "Not Found" for repos the requester does not own, so a chart you saw last quarter may not load today. Access rules like this change, so check the GitHub changelog before each quarter's review instead of relying on what this table says.
Use the other sources for narrow jobs only:
- GH Archive is updated every hour. It supplies the weekly event pull and the stargazer identities.
- The GitHub REST API allows 60 requests per hour without authentication and 5,000 per hour with it, and some limits cannot go above 12,500 per hour. Use it in step five for license, org and per-repo metadata on survivors, never for the whole universe.
- Dependents, meaning other repos or packages that declare this one as a dependency, are a tiebreaker you check by hand on survivors. No published study measures how far new dependents lead a funding round, so I do not score on them.
What the event pull costs and how to cap it
A seven-day pull of daily tables with a narrow set of columns costs roughly a gigabyte, far inside the free monthly tier. The cost traps are the first backfill, monthly tables, and the belief that LIMIT cuts the bill.
| Scope | Bytes processed |
|---|---|
| 1 repo, 1 day table | 157.69 MB |
| 1 repo, 1 month table | 5.08 GB |
| Full month of data (older estimate) | ~225 GB |
| Free tier per month | 1 TB |
Three things follow from these numbers. First, BigQuery bills for the tables and columns a query reads, not for how many repos you filter down to. A published single-repo, single-day example already processed 157.69 MB, so seven daily tables come to about 1.1 GB whether your universe has 20 repos or 2,000. Second, about four full-month scans fit in the free tier. A 180-day backfill to fill both 90-day windows would pass 1 TB at the older full-month figure, so do the backfill one month at a time with the narrowest columns you can, then append weekly. Third, LIMIT does not reduce cost on tables that are not clustered, and GH Archive is not.
On price, an older practitioner write-up quotes $5 per TB. I have not verified the current price, so check the BigQuery pricing page and treat that figure as a placeholder.
SELECT
repo.id AS repo_id,
repo.name AS repo_name,
type AS event_type,
actor.id AS actor_id,
actor.login AS actor_login,
created_at
FROM `githubarchive.day.*`
WHERE _TABLE_SUFFIX BETWEEN
FORMAT_DATE('%Y%m%d', DATE_SUB(CURRENT_DATE(), INTERVAL 7 DAY))
AND FORMAT_DATE('%Y%m%d', DATE_SUB(CURRENT_DATE(), INTERVAL 1 DAY))
AND type IN ('WatchEvent', 'PushEvent', 'PullRequestEvent', 'ForkEvent')
AND repo.id IN (SELECT repo_id FROM `my_project.oss_screen.universe`)Swap in your own project and dataset names, dry-run it to read the byte estimate, and set a byte cap on the job before running.
If you want commits counted per push rather than push events counted, add the payload column. Dry-run the query with and without it and write down the difference, because that column decides your byte cap.
How to score acceleration and strip fake stars
Score each repo on the change in rate over the latest 90 days compared with the 90 days before, and keep it only if stars, distinct outside human contributors and commits all rise together. Then drop any repo with a month where more than 10% of new stars match a fake-star pattern, unless it is a launch burst that keeps growing past 60 days.
Which window, and why
Published sources disagree, so here is the choice and the reason. Runa Capital's ROSS Index, the main public benchmark, ranks repos with 1,000+ stars by their maximum star growth over a 90-day window ending in the target quarter. Its older formula used annualized growth, AGR = (value now / value one quarter ago)^4 - 1. The vendor source uses 14-day windows for commits and 30-day windows for contributors.
I use the ROSS 90-day window as the main score, measured against the 90 days before it, and add a 30-day concentration check. Bought stars lose their effect in under two months, so a 90-day window still contains a bought bump at full strength. The window alone does not filter fakes out. The concentration check is what does it: if most of a repo's 90-day star gain landed in a single 30-day block while contributors and commits stayed flat, treat it as a spike, not momentum. "Most" is my working cut at more than half. Adjust it against your own results.
What each signal proves, and how it lies
| Signal | What it proves | How it lies |
|---|---|---|
| Stars (unique WatchEvent actors) | Attention is growing | One viral post, or bought stars |
| Distinct outside human contributors | Outsiders are investing effort | Bot accounts counted as people |
| Change in commit rate | The team is building faster | AI repos commit constantly at every stage, and bots inflate counts |
| New dependents (manual) | Other projects rely on it | No published lead time, so it is a tiebreaker only |
"Outside" contributors means people who are not already core maintainers. For a sense of scale: according to Runa Capital's annual report, the 50 companies in the annual ROSS Index together added 12K contributors and 494K stars over the following year. Contributors are scarcer than stars and harder to fake, which is why I weight them.
The fake-star filter
The peer-reviewed reference is StarScout, published at ICSE. Run over more than five years of GitHub event data, it flagged about six million suspected fake stars across 26,254 repositories before post-processing. Of those, 1.06 million matched the low-activity pattern and 4.93 million matched the lockstep pattern. In this guide, the low-activity pattern means accounts that do almost nothing except star repos, and the lockstep pattern means groups of accounts that star the same repos at nearly the same times. After post-processing, 18,617 repositories had fake-star campaigns. StarScout's code is public, and it works on the actor IDs that GH Archive WatchEvents carry.
The facts that set the thresholds:
- StarScout counts a repo as running a campaign when the flagged share of a month's stars exceeds 10%.
- A separate study used 5% as its display threshold, treating lower rates as normal background noise on GitHub.
- Fake-star sellers often sell a minimum of 50 stars, so a small repo can show a high flagged share from a single purchase.
- Fake stars mostly promote AI/LLM, blockchain, tool or application, and tutorial or demo repos. Expect more flags in exactly the AI categories most funds are screening.
- Fake stars only boost a repo for less than about two months.
So the bands are: below 5% flagged in every month passes. Between 5% and 10% passes, with the number written on the entry. Above 10% in any month means the repo is dropped, with one exception, covered in the failure modes below: launch bursts.
Momentum versus purchased attention
How to resolve a repo to a company and named founders
Start from the repo and work outwards: org domain, commit author email domains, license holder, package registry owner, and only then a people search. Founders describe their product rather than their license, so searching headlines for "open source" finds very few of them and lets in non-founders.
The numbers from Refolk's index make the case. Searching profiles that have a Founder or Co-Founder title and "open source" in the headline gives a small pool, and narrowing to a thesis keyword nearly wipes it out.
| Query | Total profiles | Non-founder titles in a sample of 25 |
|---|---|---|
| US, "open source" | 268 | At least 4 (16%+) |
| Germany, "open source" | 65 | At least 6 (24%+) |
| US, "open source LLM" | 4 | 1 of 4 |
The non-founders who slip through have real titles that sound close enough: Field CTO, VP OSS and Platform Engineering Lead in the US sample, and Staff Engineer, Software Engineer and AI Program Manager in the German one. The German sample also shows why repo names mislead. It includes deepset, the company behind Haystack, and dltHub. The repo a fund tracks and the company a founder lists on their profile often have different names.
Evidence layers for repo-to-founder resolution
- Org and website domainWho claims the project publicly
- Commit author email domainsWhich company's staff do the work
- License holder and registry ownerWho holds the copyright and publishes the package
- People search, then domain matchNamed founders whose commit emails match the company domain
No public source documents this step as a single method, so here is the order I use. Read the org profile and website for a domain. Pull commit author emails from the top committers in the scoring window and group them by domain. Read the LICENSE file for the copyright holder. Check who owns the package on its registry. When two or more layers point to the same domain, you have your entity.
Only then search for people, and search for the committers you already know about, not for a title. This is where Refolk saves time: you describe the people in plain English and get back matches across GitHub, LinkedIn and the open web, already linked to their commit identities, instead of joining a commit log to a separate people database by hand.
Confirm every founder the same way: the commit author email domain matches the company domain. Without that match, a founder title is a claim, not evidence. You are done when each survivor has a named legal entity and one to three named founders.
How to check funded status before a company enters the queue
Run the funded-status check on every candidate, not just the ones that look unfunded, and use at least three sources. A missing Form D is weak evidence, because most private raises never get one.
Form D is the notice filed with the SEC after a private securities sale. It must be filed within 15 days after the first sale, and the first sale counts from the moment the first investor is irrevocably committed by contract. A filing gives you a date. A missing filing gives you very little. A Lehigh University study found that more than half of private companies that raised money never filed a Form D, and Form D does not cover companies raising outside the US at all.
Check these sources, in this order:
- Form D on EDGAR, searched under the legal entity name from step five, not the repo name.
- The company's own announcements: blog, changelog, careers page, press page.
- The founders' profiles, where funded founders often name their investors or the round. Refolk returns founder profiles together with company history, so you can run this check with no extra lookup.
- Your CRM, for anyone a colleague already owns.
Watch the activity curve as well. GitHub involvement drops in the twelve months after a round, so a repo that peaked and is now slowing should be treated as probably funded until you prove otherwise.
Log a drop reason for every removal so the weekly numbers can be audited later:
- Funded, with the source and date.
- In CRM, with the owner.
- No entity: no company could be resolved, so the project is likely a public good or a side project.
- Founder unconfirmed: no commit email domain matched the company domain.
How the screen goes wrong
Most failures produce false positives: repos that look like momentum and are really noise, hype, or already funded. Each failure below has a specific check, and the checks are cheaper than one wasted partner meeting.
A repo that shows up on every ranking has usually been seen by every fund already.
- Ranking on stars alone. The vendor source reports that stars on their own were found to be non-predictive. What it looks like: one viral post drives a single high-scoring week. Check: require contributors and commit rate to rise too.
- Launch bursts mistaken for bought stars. In a multi-agent study, AutoGPT's flagged stars were concentrated around its launch, consistent with launch-amplified activity rather than purchase. What it looks like: an organic hype month crosses 10% flagged. Check: see whether the repo's growth persists past 60 days. Bought stars stop boosting a repo inside two months, while real launches keep pulling in contributors.
- AI repos that commit constantly. The commit signal is noisy for AI-pure companies, which commit heavily at every stage. What it looks like: a high commit level with a flat rate. Check: score the change in commit rate, never the level.
- Bot commits and bot contributors. Automated accounts inflate both commit counts and contributor counts. What it looks like: a contributor jump made of accounts that only open dependency or formatting pull requests. Check: count distinct human actors and exclude bot accounts before scoring.
- Treating a missing Form D as proof of no funding. More than half of private raises are never filed. What it looks like: a "clean" candidate who announced a seed round on the company blog. Check: add announcements, founder profiles and the CRM.
- Finding founders from headlines. At least 16 to 24% of title-filtered samples in Refolk's index are not founders. What it looks like: a Field CTO or VP OSS queued as the founder. Check: match the commit author email domain to the company domain.
- Already-funded leakage. Benchmarks list companies with up to $100M raised. What it looks like: the screen's best candidate turns out to be a well-known funded company. Check: run step six on every candidate, never only on the ones that look unfunded.
One limit applies to the whole screen. Nobody has published how many repos in one category clear thresholds like these each week, or what share of fast-growing AI or crypto repos get flagged. Your first few weeks are your baseline. If a category produces zero survivors for a month, loosen the concentration cut before you loosen the fake-star cut.
How to keep the screen current and write the queue
Three things drift: GitHub's access rules, query cost, and how well your thresholds predict rounds. Re-check the first two every quarter, and calibrate the third against your own record of which queued teams later raised.
- Access rules. Read the GitHub changelog for API and UI restrictions. Star data access has already changed once, and the fake-star filter depends on actor IDs staying available in GH Archive.
- Cost. Compare each week's dry-run estimate with your log, and check the BigQuery pricing page before you budget a backfill.
- Calibration. For every queued entry, record whether a round was later announced and on what date. After a few quarters, work out your own lead time. Published lead times describe first rounds in general, not this screen.
- Universe drift. Add new topics and seed orgs as the category's vocabulary changes, and save them as repo IDs.
Write each entry in a fixed format so a partner can act on it without reopening your research.
Company (legal entity): Repo(s) and repo IDs: Founders (1 to 3): Name | role | profile URL | commit email domain match (Y/N) Dated trigger event: YYYY-MM-DD | what happened | URL 90-day change vs prior 90: stars __% | outside human contributors __% | commit rate __% Worst-month fake-star flagged share: __% (month) Funded-status check: Form D (none / found, date) | announcements | founder profiles | CRM owner Why now, one sentence: Sources: Screened on: YYYY-MM-DD
One block per company. Leave no field blank; write "none found" with the date checked instead.
Before the queue goes out
- The universe is saved as repo IDs with written inclusion rules, and this week's new matches are added.
- The event pull was dry-run and ran under a set byte cap.
- Every queued repo shows a rise in stars, outside human contributors, and commit rate on the 90-day comparison.
- Bot accounts were excluded before contributors and commits were counted.
- Every queued repo carries a worst-month flagged share, and none exceeds 10% without a launch-burst note showing growth past 60 days.
- Every entry names a legal entity and 1 to 3 founders confirmed by commit email domain match.
- The funded-status check ran on every candidate using Form D, announcements, founder profiles, and the CRM.
- Every removed candidate has a logged drop reason.
- Every entry has a dated trigger event and source URLs.
- The queue holds 5 to 15 entries and the week's run took under three hours.
Questions practitioners ask
Can I still use the GitHub stargazers API to check a repo for fake stars?
Not for repos you do not administer. GitHub has limited the stargazer list endpoint to admins and collaborators, and third-party charting tools now return "Not Found" for repos the requester does not own. The newer star history endpoint returns counts but no identities, and fake-star checks need identities. Use WatchEvent rows from GH Archive, which still carry the actor ID of every stargazer, and run the low-activity and lockstep checks on those.
What time window should I use to measure GitHub star growth for investing?
Compare the latest 90 days with the 90 days before, and also check how much of the gain landed in a single 30-day block. Runa Capital's ROSS Index uses a 90-day maximum-growth window. Fake stars lose their effect in under two months, so growth that shows up in one 30-day block with no sustained 90-day rise is suspect. Score the change in rate for stars, outside contributors and commits together, never stars alone.
How early do GitHub signals show up before a startup raises its first round?
The peer-reviewed answer is months, not weeks. Roche et al., studying 160,065 US startups, found that activity speeds up in the twelve months before a first round, and that funded startups start their GitHub activity five months before the round on average. One vendor claims a 3 to 6 week lead before Series A, but its method is unpublished, so treat that figure as unverified.
How much does it cost to run a weekly GH Archive screen in BigQuery?
At a narrow column set, a weekly pull is small. A published single-day example processed 157.69 MB, so seven daily tables come to roughly 1.1 GB, well inside a 1 TB monthly free tier. The expensive part is the first backfill: a full month once scanned about 225 GB. The current per-TB price is not verified here, so dry-run every query, and remember that LIMIT does not cut the bytes billed.
Is a missing Form D proof that an open-source startup has not raised?
No. A Lehigh University study found that more than half of private companies that raised money never filed a Form D, and Form D only covers US offerings. Treat a missing filing as weak evidence. Also check the company's own announcements, the founders' profiles and your CRM, and remember that GitHub activity that slows down after a burst often means a round has already closed.
Why not just search for founders with open source in their headline?
Because the pool is tiny and noisy. In Refolk's index, 268 US profiles have a founder title and "open source" in the headline, and only 4 also mention "open source LLM". At least 16 to 24% of title-filtered samples are not founders. Founders describe the product rather than the license, so start from the repo's domains, commit emails and registry owner, then search for people.
Try it on the search you came here for
Stop building boolean strings. Just describe the person.
Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.
01Describe them
One plain sentence. Role, city, stack, stage, whatever matters to you.
02I read the web live
GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.
03You read the shortlist
Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.
- Staff backend engineers in NYC who shipped Rust in production
- Series A fintechs in SF under 50 people, growing headcount this year
- Maintainers of fast-growing Rust web frameworks on GitHub
- No boolean, no filters, no seat to buy. One box.
- Read at search time, so a profile updated yesterday counts today.
- Every step visible as it runs, every name with its reason.
500 free credits on sign-up. No card, no demo call. See real searches.