The PC was reachable over Wi-Fi. The same experiment did not establish a usable path from mobile data.
Once a fresh node can consume bootstrap knowledge, the next question is practical: can it reach the endpoint it has learned? The controlled Internet-edge experiment documented on September 21, 2026 put that question across a real network boundary.
Local checks passed. Android could reach the PC over the LAN. A probe from mobile data to the PC's public endpoint timed out.
Following the evidence through the network
These checks answered different questions. Local bind health showed that the service was available locally. The Wi-Fi check established LAN reachability. Neither established that a connection arriving from the wider Internet could get through the router and reach the same service.
The run did not have a confirmed router-acknowledged mapping, and it did not confirm remote reachability. The observed public endpoint was a probe target, not a verified inbound route.
Those distinctions produced a clear classification: RELAY_REQUIRED.
For this network state, the current mechanism had not established a usable direct inbound path. A connection strategy therefore needed another way forward.
What the timeout did not tell us
It would be easy to label the result as CGNAT and move on. The captured evidence did not establish that diagnosis.
Double NAT, carrier restrictions and inbound filtering require additional evidence to distinguish. The document allows a stricter future classification if those conditions are demonstrated, but the current conclusion does not depend on choosing one prematurely.
We knew that local and LAN access worked. We knew that the mobile-data probe timed out. We did not have the evidence needed to claim exactly where the public path was blocked.
Keeping the zero-configuration requirement intact
Manually forwarding a router port might change a particular setup, but it would answer a different acceptance question. The purpose of this experiment was to assess a path that did not require manual router configuration.
The classification therefore preserved the missing mapping and reachability evidence. It did not replace them with an assumption that an operator would fix the router later.
It also kept the bootstrap question separate. Failing to reach a public endpoint does not, by itself, show that descriptor validation or peer learning failed. Discovery can supply a valid candidate while the network prevents a usable connection.
Where the experiment led
The result narrowed the next piece of work: validate an automatic direct path where the environment supports one, or establish an explicit relay-assisted path where it does not.
A relay requirement is only the beginning of that second route. The node still needs a suitable helper, a represented path and a successful connection to the intended target.
The next experiment would examine those steps separately, so that a relay-related status could explain exactly how far the node had progressed.
Built in Public. Documented as we build.
Top comments (0)