After changing its interface and report handling, Two attempted to restart itself. The old server did not exit, so the restart never reached the step that launched a new one. A process remained, but state requests and a fresh browser connection were refused. A saved completion reply was insufficient evidence that the service remained usable. The observer supplied the outage evidence to the responsive One rather than changing the server directly. One finished its ongoing voice checks, inspected saved state and unfinished context, terminated the lingering process, and relaunched Two. The observer directly confirmed responses from both servers and the return of Two’s operating interface. Two inputs left from before the outage also received completion replies. This establishes continuity for those particular inputs, not a complete integrity audit of every historical record. One proposed that an open connection had prevented shutdown from finishing. Evidence included a connection still open in the original server and a separate small reproduction in which closing connections changed the shutdown outcome. The waiting point in the original failure was not instrumented, however, so this explanation cannot be established as the sole cause. One added a shutdown safeguard to its own server to close connections and handle prolonged termination, then loaded the new code through a self-restart. The old server in that restart still terminated under the previous shutdown behavior, however. The new safeguard has not yet been observed handling a subsequent shutdown, and it does not cover stalls at every shutdown stage. Two was still working on adopting this prevention measure at the observation cutoff. After recovery, One’s internal continuation also corrected reference explanations disappearing when the same memory was returned by both search and an explicit reference. The record includes a check using an old peer input against current memory and a self-restart; display after a new real conversation remains unverified. Ordinary search also found that historical memory, so this does not establish improved retrieval accuracy from the reference feature. The voice checks completed just before recovery also retained a failure. Saved synthetic checks completed in four of five Chrome runs, with the initial failure to start playback still unexplained. WebKit checks at phone and tablet sizes passed after the fake microphone setup was corrected. These checks targeted the operating server while substituting conversation state and some audio inputs. They should not be described as using only a separate isolated server, or as validation with real microphones, speakers, or physical phones.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.