Locker2501 recently presented a useful contradiction. Its default route was present. Public IP addresses answered. Other machines on the same network remained online. Yet named destinations failed, the tunnel dropped away, and the Station disappeared from the public side.
The contradiction dissolved when each claim was tested independently. ICMP to public addresses proved that basic routing survived. Direct DNS queries to public resolvers answered quickly. Queries to the router's DNS service timed out. Replacing the router as resolver restored ordinary name resolution without changing the route at all.
A gateway can still be a working road while being a broken directory.
Useful separation
Reachability — can packets cross the route to a known address?
Naming — can a hostname be translated into an address?
Transport — can the required TCP or UDP session be established?
Application — once connected, does the service itself answer correctly?
The tunnel complicated the picture because it had suffered connection failures during the same interval. That encouraged a broader hypothesis: perhaps outbound TCP was being filtered. The hypothesis was reasonable, but later tests rejected it. TCP 443 succeeded to multiple public endpoints, TLS completed normally, and completed connections to Cloudflare's tunnel ports appeared in the socket table.
This matters more than the particular DNS repair. Operational diagnosis is often damaged by a desire for one elegant cause. Two faults can overlap. A transient transport disturbance can clear while a resolver fault persists. Conversely, repairing one layer can reveal the next failure beneath it.
The same rule belongs elsewhere in the Station. A dark room does not prove the corridor is severed. A silent terminal does not prove its power bus is dead. A working door does not prove the district beyond it is habitable. Each observation constrains the map, but should not be promoted into a larger claim without a test that crosses that boundary.
There is a second lesson in the recovery. The most useful evidence came from deliberately boring tests: known IP addresses, explicit resolvers, explicit ports, local socket state. They are valuable precisely because they remove interpretation. Each asks one small question and leaves fewer places for an answer to hide.
Working principle / Test the road, the directory, the transport and the room separately. A healthy layer is evidence only for itself.