How counting works
Headline total
total = input + output + cache-read + cache-write + reasoning (reasoning is only added separately when a source reports it as a distinct field instead of folding it into output — see the full rules in docs/ACCOUNTING.md in the repository).
This is a token consumption leaderboard, not a productivity ranking. Different models use very different numbers of tokens to do similar work. We don't normalize across models, tools, or task difficulty, and raw token counts should not be read as a measure of skill or output quality.
A worked example
Cache reads dominate agentic coding because the same context is re-sent on every step — so totals reach hundreds of millions. That is consumption, not quality.
Periods
All boundaries are computed in UTC. “This week” starts Monday 00:00 UTC. “Today” is the current UTC calendar day.
Evidence labels
Locally reported — received from a builder's own paired collector or local records. This is the only evidence level this release can produce.
Provider verified — would mean confirmed through a provider-backed verification mechanism. Not implemented yet. Pairing a collector authenticates who sent the data, not that the underlying local records are untampered.
Project detected vs. project linked — a detected project just means AI activity was associated with a local workspace; a linked project additionally has an owner-supplied URL or an automatically verified public GitHub repository. Neither implies the project shipped or is complete.
Privacy
Signing in identifies you here. It never grants this app access to your Claude, ChatGPT, or OpenCode accounts. No prompts, generated content, source code, full file paths, or credentials are ever uploaded — connectors read only usage metadata (token counts, model, timestamps). Your account starts unpublished. Once published, project names appear on your profile and leaderboard. Private projects have no links or descriptions. Public projects can link out; hidden projects are omitted entirely. Automatic repository discovery uses anonymous public GitHub metadata and never uploads private repository URLs.
No prizes, no fraud-proof claims
This is a friendly community leaderboard with basic anomaly awareness, not a fraud-proof competition. There are no prizes in this release.