Notes.
Field notes from the lab, on the architecture, the data, and the work itself. Dated; marking what runs today and what is still coming.
- Pay for verified work, not tokens The honest unit of a coding agent is the change that works, not the tokens it spent getting there.
- Standing on the frontier We build on the frontier models, not against them. The work is the efficient, specialised layer they make possible.
- The unit of account A token is cheap. An answer you cannot accept is not. The cost that matters is per accepted outcome, and it includes the tail.
- When inference has to pay for itself Metered inference already clears its cost. The reckoning is not cheaper tokens; it is each deployment having to show it was worth running.
- Collecting the right data now We have no models yet. We are recording the data that will train them, at the resolution you cannot add back later.
- Why two models fail differently Every model has one built-in way of failing. That is the opening, not the obstacle.