Where humans are involved
This system automates verification, not judgement. Ten points in the loop involve a person, and each one has an owner, a trigger, and a clock. They are listed here because "no one in the middle" was a claim v0.1 made twice and could not support — and because a human step you did not plan for is indistinguishable from an outage.
Not automation with exceptions — automation around a small number of decisions that should not be automated.
Plus, on the partner side
When this webhook fires, no amount of backing off will resolve it. Queue the affected object and surface it to an operator. Integrations that treat it as a transient failure retry for 24 hours and then give up on something that was always going to take a day.
Signed telemetry takes minutes. Reaching investment_ready takes weeks, and most of that is steps 1, 2, 5, and the legal attestation — none of which move faster because your integration is finished. Start them in parallel.
Downgrades fire without approval, immediately, and freeze new issuance on their own. A person is involved only in the correction path, and only after the protective action has already taken effect. That ordering is the point: protection first, judgement second.