Most of what happens on the Gonka chain, most of the time, is the inference lifecycle — not proof-of-compute ceremony. It shows up on chain as a pair of message types:

• Start inference — a user request enters the system. Because the actual cost of an AI completion isn't known until it's generated, the participant's API node checks the user can cover the maximum possible cost (based on the request's max-completion-tokens setting) and locks that amount in escrow on chain.
• Finish inference — the completion is done. Actual token counts (prompt plus real completion tokens) are recorded, the provider is paid from escrow for exactly the work performed, and any unused escrow is refunded to the user immediately.

The fee itself is calculated as:

Final fee = (prompt tokens + actual completion tokens) × units-of-compute-per-token × unit-of-compute price

Both the per-model units-of-compute-per-token figure and the unit-of-compute price (a weighted median of participant proposals, updated each epoch) feed into that formula — see Gonka Tokenomics for the full pricing model.

Inference itself is also validated, at a rate that scales with a participant's request volume: high-volume participants get sampled at a minimum rate, while low-volume participants (100 or fewer requests) have every single task checked.

Scroll the transaction feed for a few minutes and you'll mostly see start-inference and finish-inference pairs — this is the network's core, everyday activity, punctuated by proof-of-compute bursts near epoch boundaries. See How to Read a Gonka Block for the full transaction-type glossary.