Skip to content
Zevqi
AI Engineering

We Audited 70 Popular Open-Source Repos for Knowledge Concentration: What Git Can and Cannot Show

Median 4.5% of files rest on one author, 41 of 70 repos have a sensitive file held by one person, and heavy AI use did not change the picture. The numbers, the method and the limits.

By Neha SinghPublished November 21, 20266 min read

We ran our free zevqi audit over the last twelve months of git history in 70 popular open-source repositories. We expected to find a lot of code that rests on one person. We mostly did not, and we found no difference between the repositories that mark many commits as AI-assisted and the ones that do not. Here are the numbers, how we got them, and what we think they do and do not mean for a company team.

The short version

  • Concentration in git was modest. In the median repository, 4.5% of files had only ever been changed by one person across three or more commits. The middle half of repositories ranged from 1.3% to 11.5%.
  • Repositories where it was high were rare. 6 of 70 had a quarter or more of their files in that state, and 1 had more than half.
  • But sensitive code did not escape. 41 of 70 repositories had at least one file in an authentication, billing, infrastructure, data or deployment area that rests on a single author.
  • AI marks were common. 53 of 70 repositories had at least one commit that says an AI tool helped. Across all 79,494 commits, 10.2% carried a mark, and 15 repositories were at 10% or more.
  • Heavy AI use did not show up as more concentration. The median single-author share was 4.4% in the 15 repositories with 10% or more AI-marked commits, and 4.5% in the other 55.

How we ran it

We picked repositories by a fixed rule, before looking at any result: public, not a fork, not archived, pushed in the last 30 days, written in a programming language (not a list, book or notebook), under about 800 MB, and not named like a tutorial or a list. We took two groups by stars, 20,000 and above and 3,000 to 9,999, and from each took 40 repositories evenly spaced down the ranking. That gave 80. We cloned each one without file contents, ran zevqi audit on its last twelve months, kept only the counts, and deleted the clone. No file names, author names or emails were stored.

73 repositories produced a result. Seven did not: six hit our time limit on very large histories and one had no commits in the window. Percentages use the 70 repositories with at least 30 source files with history. The median repository had 43 active contributors over the year. Six repositories have 5,000 or more files, and for those the audit reads the 5,000 most-changed.

The rule for “one author” is deliberately strict: exactly one person has changed the file, over three or more commits in the year. Most files in any codebase change once or twice a year, so this counts a small, active slice of the code, not every file nobody else has touched.

By team size

The repositories where a few people do most of the work are the ones closest to a company team, so we split by active contributors:

Active contributorsReposMedian single-author share of filesRepos at 25% or more
Under 10110.9%0
10 to 49264.3% (top quarter above 19.9%)5
50 or more335.2% (top quarter above 7.4%)1

The spread was widest in the 10 to 49 range, which is roughly the size of the engineering teams we talk to. Eleven and 26 are small groups, so treat this as a pointer, not a finding.

What this does and does not tell you

It is not your company. Popular open-source projects have many contributors, public review and a culture of shared ownership. A five-person product team that ships fast with agents can look very different. Use this as a baseline, then run the audit on your own repository.

AI-assisted is a minimum. We count a commit only when it says so, through a co-author line or a generated-with note. Many tools leave no mark, so the real share is higher, and our comparison between high-AI and low-AI repositories is really a comparison between repositories that mark their commits and ones that do not.

Git shows who changed a file, not who understands it. This is the most important limit, and it is the reason the headline result is not reassuring. An agent can change a dozen files in a minute, and the person whose name is on the commit may never have read them. A repository can look evenly shared in git and still have code that nobody could explain. Git cannot see that, because it records changes and not comprehension.

What we took from it

If concentration in git is usually modest, the risk to look for is not only “one person wrote this”. It is code that changed and that nobody has demonstrably read, plus the smaller set of sensitive files that do depend on one person. The first needs a record of who looked at which version of a file, and the second needs a way to see what would be left if someone went. That is what Zevqi records and shows, on top of the git view you can get for free.

Run it on your repository

From inside any git repository:

npx zevqi audit --html

It reads git history on your machine, sends nothing, and writes a one-page report you can share. The rules are printed in the report, and every flag says why it was raised. Details are on the audit page. If you would like the per-repository counts behind this post, email us and we will send them.

Frequently asked questions

Which repositories did you analyse?

We chose them by a published rule, not by name, and we report only totals. We do not publish per-repository figures, because a single-author file is a normal state for many healthy projects and we do not want to single anyone out.

Does a single-author file mean a project is at risk?

Not by itself. It means the history gives you no second person to ask. It matters most when the file is sensitive, when its author has gone quiet, or when it is hard to change safely.

Why not measure understanding directly?

Nothing can prove it. The practical substitute is evidence that a person looked at the current version of a file and said whether they followed it, which stops counting when the file changes. See how to measure understanding of AI-written code.

Related: bus factor in the AI era.

Written by Neha Singh

Neha Singh is the founder of Zevqi. She builds and tests the tools on this site and writes these guides from hands-on use. Follow on X (@heynehasingh) or read more about Zevqi.

Try it now

Ready to put this into practice? Zevqi has the tools built in — free, no signup.