A narrow signal, measured reproducibly.

Operational definition

Considered commits are commits in the fixed calendar window after merge commits and routine CI or dependency bots are removed. The primary ratio is unique agent-attributed commits divided by considered commits.

agent-share floor = agent-attributed commits / considered commits

Because squash merges, rebases, and manual commits can remove or omit signatures, the result is a lower bound—not an estimate of AI-written code.

Detector validation

The S1 study independently labelled a stratified blind sample of 600 commits and adjudicated 19 disagreements. The publication gate required precision of at least 97%.

Precision

99.45%

Wilson 95% interval

97.77%–99.86%

Weighted recall

72.86%

Raw agreement

96.83%

Cohen’s kappa

0.938

Gate

pass

Rules and safeguards

The versioned detector matches supported self-declarations in author identities, commit trailers, and agent-specific metadata. Validation explicitly checks documentation examples, reverts quoting an earlier message, and repositories with signature-bearing templates. Generic prose or code style is never classified.

The full ruleset is identified by sha256:b2e8889c66f72c18a61839f0bf1a9f77b5481ba2def044dba33c797b2f2bdcae. A result without that hash and its measurement window is not comparison-compatible.

Interpretation limits

1. Zero means no supported signature was observed; it does not mean no AI was used.
2. Commits are units of history, not lines of code, effort, quality, or productivity.
3. The stars-ranked sample is descriptive of this corpus, not representative of all software.
4. Low-activity observations below 100 considered commits remain downloadable but are excluded from rankings.
5. Cross-ecosystem differences are descriptive and must not be interpreted causally.

Corrections and privacy

Repository identities, redirects, default branches, and resolved HEAD SHAs are published so data-quality reports can be checked. Corrections are documented rather than silently overwritten. Read the tagged correction log ↗.

Contributor identities are used only to compute aggregate concentration. Names and email addresses are not exported, and qmmit must not be used to rank, surveil, or evaluate individual developers.

On-demand scan lifecycle

An unknown repository never receives an ambiguous indefinite spinner. Worker responses map to explicit states, and only temporary failures offer a retry action.