Methodology

Methodology version 2026-09-13. Read as Markdown

DevImprint measures one narrow thing: commits published in public repositories that carry a detectable signature from an AI coding tool. It is not a measure of productivity, effort, seniority, or code quality, and it cannot see private work.

What counts as an AI-assisted commit

DevImprint associates evidence in public repository history with AI coding tools. This is an observational classification: it does not verify authorship, establish the amount of assistance, or inspect private work. More than one tool can be associated with the same commit, so tool populations can overlap.

Exact matching rules, attribution identifiers and detection examples are not published, to limit manipulation of the measurement. Public documentation covers definitions, publication scope, time windows, provenance, coverage and limitations. Tool pages report observations within that scope.

False positives and missed activity are possible. A detected association is evidence, not proof; absence of an association does not establish that AI was not used. Report suspected errors with the affected profile and snapshot date.

The ranking window

Rank is based on the count of AI-assisted commits authored in the trailing 30 days, currently the window beginning 2026-08-15. The downloadable data preserves both global rank and separate ranks for people, tools, and unresolved identities, so a bot can never displace a person in the default people view. Each category has its own URL, and its ranks remain stable across pagination. Lifetime totals are shown alongside but do not affect rank, and ties are broken toward more recent activity.

Who appears

The upstream pipeline labels every row as a person, tool, or unresolved identity before publication. People are the default view; the other categories remain available through the filters and downloadable data because their activity is still part of the measurement.

• Email-shaped identities. When commit history yields only an address and no resolved account, the address is replaced upstream with a stable opaque handle. The address itself never enters the public document.
• Tools and automation. Bot accounts are the thing being measured, not competitors to people. A commit attributed to an automated identity counts in the automation category and toward the repository, rather than toward a person's rank.

The site builds 846 profile pages, one for each person in the published ranking. Tools and unresolved identities have no profile: there is nothing behind the name to show.

Unresolved identities

Not every commit can be traced to an account. When history yields only an address, that row is ranked and counted but published as an opaque handle such as unresolved-a1b2c3d4e5f6, never as the address itself. The address is real data, and republishing it on a ranked page would disclose something its owner never chose to display, so it stops at the pipeline. The handle is stable, so the same identity keeps the same label between snapshots and you can watch it resolve.

The published subset contains 846 people, 19 tools and agents, and 135 unresolved identities. These counts describe the downloadable subset, not the complete ranked population. Each group has its own ranking page.

They are also not who you would guess. Measuring contributor linkage across the ranked repositories, an identity is about twenty times less likely to be unresolvable on a repository with a single contributor than on one with several. Unresolved identities are concentrated in large, long-lived, many-contributor repositories, which is the signature of imported and vendored history, merged forks, and commits authored under addresses never registered to an account. Most of them are not developers we could ask; they are history the repository inherited.

What the method cannot see

• Private and internal repositories. Nothing outside public history is read.
• Work where the developer turned tool attribution off, which many people do.
• Changes to repository history that affect the available evidence.
• Tools that leave no signature at all, including most inline completion.
• Repositories that have not yet been scanned. See coverage.

These gaps can lower observed counts, while attribution errors can also inflate them. Treat the figures as observations with limitations, never as a verified total or lower bound.

Provenance

Each published snapshot carries the digest of the artifact it was produced from, the time it was generated, and the coverage that stood at that moment. The current snapshot isda8f0f66a1eb4f0f, generated 13 Sept 2026, 03:27 UTC. The same document that backs these pages is downloadable at /leaderboard.json.

Tools detected in the current snapshot: Aider, Blackbox, Claude Code, Codeium, Codestral, Codex, Cody, Continue, Copilot, Cubic, Cursor, Devin, Jules, Netlify Coding, Openhands, Replit, Sweep, Windsurf.

Opting out

Anyone can ask to be removed. A removal takes effect on the next build and covers the leaderboard, the profile page, the downloadable document, and the social card; the profile URL then returns a permanent “gone” response rather than a redirect. Aggregate counts retain the commits without the identity attached.

A self-serve control arrives with profile claiming. Until then, removal requests go to[email protected] from an address or account that can be matched to the profile. No reason is needed and none will be asked for. What to send, and what happens next →

Removal also deletes everything held privately about you — see the privacy notice, which covers the information that supports shortlists and never appears on this site.

Corrections

If a number here is wrong, the fix and the reason are published rather than quietly applied. Attribution errors are the most likely kind: incorrect identity linkage or changes in attribution conventions. Reports are welcome at the same address.