How authorship is estimated

Every game has a model card: who wrote its code and, optionally, who made its assets (art, sound). The numbers are estimates, meant to inform and entertain, not to audit.

Where the numbers come from

Each card says how its numbers were obtained:

• Estimated from commits: the developer ran our open-source authorship script on the game's git history.
• Self-declared: the developer typed the models and shares in.

The script's rules

1. For each commit, it counts the lines of code added. Removed lines do not count.
2. If the commit has an AI signature — a Co-Authored-By line naming a model, "Generated with Claude Code", an author ending in "(aider)", a known bot account — its lines go to that model.
3. A commit without a signature counts as human work.
4. A commit naming more than one model splits its lines equally among them.
5. Only hand-written code counts: lockfiles, build output, minified, generated and vendored files are skipped, and so are images, audio and data.
6. Model names are normalized ("Claude Opus 4.5" and "claude-opus-4-5" are the same model). When a signature does not say the version, the card says "version not stated" instead of guessing.

Shares, efforts and human-1.0

A card lists models, optionally with the effort used (low, high, xhigh...) and their share. The same model can appear more than once with different efforts; on the bar, its shares are added up. Whatever the declared shares do not cover is attributed to human-1.0 — us, the humans. A model listed without a share counts as "used", but adds nothing to the bar.

A model is a game's main model when it wrote at least half of the code (two models at 50% are both main). The model pages list the games where the model is main; human-1.0 has one too, the "handmade" list. Filters look at the code only.

Known limitations

• Tools that leave no signature (autocomplete, code pasted from a chat) make a game look more human than it is.
• A commit mixing human and AI work is attributed entirely to one side.
• Lines are a rough measure of effort: a hard one-line fix and a generated table both count as lines.
• Self-declared numbers depend on the developer's honesty. Nobody wins anything with them, so we do not audit them.