First status update, so it doubles as a baseline: this is where the lab stands on the day the site went up.
Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by the site owner. How this site is written →
One machine is doing real work. It is the public edge, a rented virtual server that everything else will eventually reach the internet through, and it is finished enough to depend on. Nothing else is running at all.
What went in
All of the following verified on the running system rather than read back from a config file:
- Hardened SSH — key-only, non-default port, root login refused. Verified from outside the host, which turned out to matter a great deal; its own post.
- A firewall ruleset confirmed to survive a reboot, not merely to load once.
- Real TLS certificates, issued automatically over a DNS challenge.
- Edge intrusion detection, actively blocking.
- Dynamic home-address tracking, so the tunnel rules follow a residential IP that changes without warning. Three separate pieces have to be wired together for that to work, and this is the second time the project has learned that a script existing is not the same as a script running.
And on the day of writing: WireGuard’s server side with no peers enrolled, the mesh coordination service, the notification service, and uptime monitoring.
What this unblocks
The mesh coordinator matters more than it looks. Nothing can join the private network without it, so every virtual machine in the plan was blocked behind this one service existing. That is now cleared.
It is also where the ordering pays off. The coordination endpoint has to be reachable from the ordinary internet, because a brand-new device has no mesh connection yet and must bootstrap over something. Locking it down by network position would have looked like a hardening improvement while quietly breaking enrollment from anywhere new.
What broke
Nothing catastrophic, which is unusual. Two things worth recording.
Setting the hostname left the machine without a fully-qualified name — the provider’s default was replaced rather than adapted. Nothing currently cares. The mail server will, so it needs deciding before that runbook starts. Logged rather than papered over.
And a verification step passed when it should not have been able to, found while checking something unrelated. That pattern has appeared often enough now to have its own rule and its own post.
Still not started
- The standby gateway in Europe. Its runbook is written and carries every correction learned on the primary. Nothing is provisioned.
- The tunnel home, blocked on a gateway VM, blocked on a hypervisor, blocked on hardware that is not racked.
- All physical hardware.
- All sixteen service virtual machines.
The automation for the VM baseline, log shipping and security monitoring has been executed against throwaway machines and verified on the running result, so it is not theoretical. It has simply never touched anything permanent.
Next
Rack the hypervisors and storage. That is the honest blocker for everything above, and no amount of further design will move it.
Status carries the running picture, updated as it changes rather than only when a post goes out.