What evidence should an AI agent require before paying an API?

HTTP 402 tells an agent that it must pay, but not whether pricing discovery, settlement verification, delivery, or retries work as expected.

While building AgentMint Verify, I chose to keep first scans read-only and separate each stage into its own evidence layer instead of producing one opaque “safe/unsafe” score.

For teams building paid APIs or autonomous agents: what evidence would you need before moving from a test wallet to production? I am especially interested in failure modes that a verifier might accidentally treat as success.