Every team using AI coding tools eventually asks the same question: why did that task cost so much?
The answer is usually not “the model was expensive.” It is context that had to be rebuilt. A retry after an ambiguous instruction. A stronger model used for a routine scan. A handoff that lost the thread. A review that caught a failure only after the work had already been paid for twice.
AI work has an economy, but most interfaces show only a total. A total is a receipt without line items. It tells you what left the account, not what produced progress.
The cheapest capable path
Routing should not mean always choosing the cheapest model. It means choosing the least expensive model capable of the work, then escalating when ambiguity, risk, or rework makes the premium rational.
That requires recording the decision: what kind of work was this, which path was chosen, what evidence came back, and did the result hold up?
Maximizing intelligence consumed is easy. Maximizing verified progress per unit of intelligence is the product.
What a useful receipt contains
A useful receipt makes input, cached context, output, reasoning, retries, model route, and verification visible together. It lets a team ask better questions: did caching help? Did the expensive route prevent rework? Are we spending Sol-level intelligence on Luna-level work?
That is the layer Cairn is building. The work stays in the agents and repositories people already use. Cairn keeps the economics and the proof beside it.