The field doing the most work in that list is the idempotency key, and it is also the easiest to fill with a UUID and forget. The post says the key prevents duplicate sends, duplicate updates and duplicate refunds [11]. An envelope cannot do that by itself. A key only stops a second write if the endpoint receiving it stores keys and refuses replays, and the post names its target systems (Gmail, CRM, billing, calendar, file storage, Slack, Discord) without saying which of them accept a caller-supplied key on the write paths agents use [12][3]. That is a per-endpoint check somebody does by hand: refund creation, stage transition, user invite, message send, one at a time. Where the answer is no, the key records an intention and the duplicate refund still lands.
Expiry is the field that changes the shape of an approval [17]. In the flow, the approval inbox is not the last gate; a state re-check sits between the approval and the action [4][2]. Those two together make an approval a lease rather than a verdict. A human said yes to a described world, and if the invoice was already credited or the lead already moved, the yes no longer refers to anything. Setups that approve in chat and fire on the click have no such gate, which is why the post's list of unanswerable questions includes whether the business state changed after approval [9].
The tension inside the schema is who classifies. Risk class is a field the requester fills in [5], with values running from low to blocked [6], while deny and escalate are outcomes the policy layer produces [7]. Classification therefore appears twice, once as a self-report and once as a decision [4]. A policy layer that trusts the self-report has handed its job to the thing it exists to constrain. The envelope already carries enough to classify independently: target system, affected record, exact action [5]. Risk class reads better as the requester's opinion, logged and compared against the outcome, than as the input to the branch.
What makes the pattern auditable is duller than the risk taxonomy. Each decision carries a reason from a fixed set: missing evidence, customer impact, money movement, public visibility, policy mismatch, stale state [8]. The envelope also names a log destination [3], so those reasons land somewhere queryable instead of in the thread where somebody clicked approve. A month of that log tells you which writes were proposed, by whom, and on what evidence, which is exactly what a list of sanctioned connectors cannot tell you [10].
