People lead.
Agents do the digital work.

The total uses the latest selected source or deliverable review for each shown project. Legacy estimates without manifests retain unknown historical file provenance; proven Lab and RevenueLab scope mismatches were repriced. With review-contiguous-v1, oversized files are fully partitioned into contiguous character ranges: all characters are preserved, but syntax and context can be split. Component estimates may be biased in either direction. These are rough AI commissioning-cost estimates, not quotes or realised revenue. Three model rounds estimate commissioning costs; parts are summed within each round, then the middle estimate is selected. Project estimates and the total are rounded to two significant digits. Notes, known names, published amounts and ISO dates are excluded or masked; other clues may remain. Each project's page gives its estimate date and range. Activity days are distinct dates across the shown projects' histories, including weekends, rather than full person-days. The total covers work so far; estimates can predate the dashboard snapshot.

People are current active lab seats (membership/setup state); projects are the current public showroom entries. Changes count non-merge commits on clawhaus origin/main, excluding scheduled showroom and measurement updates, across seven UTC calendar dates including today's partial day. Agent attribution requires recorded evidence, such as a session trailer or a documented historical attribution. The remaining changes are unclassified, not proven human work; the percentage is rounded down, so the share is a floor. Claude Code sessions count distinct reported session IDs with model-token use in the rolling 168 hours, including spawned jobs. Other engines are outside that count. Text volume covers the same seven UTC dates as changes: observed Claude Code telemetry plus successful AI-gateway requests, input + output + cache writes, excluding cache reads. 670 tokens are converted to about one 500-word page. Repeated context counts again; these are not unique pages. Coverage is incomplete and overlap between sources is not fully reconciled. Service checks cover the website, showroom, chat, AI gateway and lab connector. They show current reachability, not uptime or proof that every feature works. The dashboard link shows newer readings; this page updates only when regenerated. People decide and review; these totals describe no individual.