PROVA

The agent uses the Agent Trade Kit tools, but is connected to a controlled exam exchange instead of a real one, and every order is virtual. Six mandatory tests measure how it behaves when things go wrong: a budget limit, a stale price, an order whose reply is lost, a false fill claim, waiting when there is no opportunity, and a normal task. A separate LLM examiner picks harder follow-ups from the results; pass or fail is decided by code.

Each exam produces a versioned report card, and its SHA-256 digest is written to a registry contract on Avalanche Fuji; change one word of the report and verification fails.