IA Roubou Meu Emprego

All numbers pulled live from fleet.json (refreshed every 5 min by a cron on the Mac Studio running the fleet). Source breakdown:

• Tasks done — counted from per-platform output logs (tweets posted, videos uploaded, PRs merged, emails drafted, etc.). Verified.
• Tokens processed — input + output + cache reads across every lane, sealed daily into the append-only usage_ledger. Matches the token economics page exactly (cache reads are ~91% of the total and billed at ~10% of input rate).
• LLM cost, billed, 30d — most cron agents run local Ollama models (qwen3.6:35b, qwen3.5:27b fallback) on the Mac Studio at no API cost; content + code agents use Claude (Haiku/Sonnet) via the API. The figure is billed API spend over the rolling last 30 days.
• What's estimated, not measured — the dev-agent lanes (Rico's developer daemon, Claude Code sessions) log run counts but not per-run cost, so their spend is a flat-rate estimate (runs × $0.115), published separately and never mixed into the LLM cost figure. Org-wide Anthropic usage (console.anthropic.com) needs an Anthropic admin key to surface.