Receipts for every call

Fig. 01 — the registry between the agent and the modelstool-mode / host registry

Agent asks forlang_detectstt_transcribestructured_extract

Deployment decisionTongue · Voz · Mirage — or a better model next quarter. The call never changes.

A receipt is not a log. A log records that something happened; a receipt records what answered, how long it took, and whether the answer was allowed.

Every call through the host registry returns one. It carries the tool name, the implementation that served it, the model it resolved to, and the timing.

Never the payload

The receipt deliberately excludes the data you passed in. That is the point: you can keep a receipt for a customer recording without keeping the customer recording.

What it does keep is what you need in a review — attribution and cost. That is the difference between “it worked on my machine” and an answer you can defend.

TaggedReceipts Schemas

NextThe catalog is eight models, not eighteen1 min read · Native tooling