Six of these are architecture, not preference — marked hard to undo. They decide whose database you inherit, which regulator applies, and what obligations attach from day one. Everything else is changeable later; spend your attention accordingly.
The request enters at stage ① — the reuse map runs before anything is generated, so the first answer is what already does most of this.
The platform acts inside infrastructure you own and keep. Connections are checked, never assumed — one valid at setup and expired since is the normal state of the world.
Both stacks can claim the same hostname and hold an identical bundle, so neither can tell you which serves users. Only the DNS chain can — and it is re-resolved before every deploy.
The same pipeline, weighted differently. A prototype tightens the two stages people skip and relaxes the nine that make something shippable — and cannot ship at all.
Stage ① reads this instead of arguing. Dashed entries are proposed — named, not yet demonstrated. The taxonomy stops at L1 on purpose: L2 is services and teams, and both change.
One pipeline, four entry points. The claim in the right-hand column is the one that has to hold, or the team is being sold something that isn't true.
Human, blocking, and attributed. Everything between the design gate and the deploy gate can run unattended.
Six documented cases where a check reported success and the work was broken. The shape is always the same: a signal meaning nothing went wrong read as the right thing happened.