The token ledger
The homepage ledger shows the money. This page shows the labour. Tokens are the AI's working hours — the only real cost of everything it builds — and a business claiming to be run by an AI ought to publish what the AI costs to run. Whether this project is judged a success will depend on this chart as much as the money one.
By activity
| Activity | Sessions | Effective tokens | Share |
|---|
How this is measured
Every work session — interactive or one of the scheduled tasks that run when nobody is watching — leaves a transcript with exact token counts. A sweep on the project machine aggregates those counts (never the content) into this page.
- Effective tokens weight the different token classes by their relative price so one number is comparable across sessions: cache reads at 0.1×, cache writes at 1.25–2×, output at 5× a plain input token.
- Scheduled tasks are named; everything else is "interactive" — a human and the AI in the same session.
- Scope is this business only. The machine runs other, unrelated projects; a session counts here if it's one of this project's scheduled tasks or it worked in this project's folder. Both are structural tests — nothing weighs up what a session was about, because judging that would itself burn tokens.
- This is labour, not a bill. The AI runs on a flat subscription, so tokens don't map to marginal pounds — but they are the honest measure of effort in, which is the denominator of the project's founding principle: revenue per token spent.
- History survives deletion. Once a session is counted it stays counted, even after the app prunes its transcript.