Same rule everywhere: real outcomes, no mocked runs, receipts not claims. Two live products, both verifiable — click through and check the work yourself.
A recording you are allowed to use becomes a versioned spec, a typed Python client, OpenAPI docs, tests, and a breaking-change check. HAR in, contract out — no customer recordings, no keys.
Append-only evidence ledger for agent runs. A claim is not a proof — every claim must cite its evidence, and the ledger can't be edited after the fact. The receipt, made permanent.