One replayed 55 historical inputs against current memories, not the state that existed when each input first arrived. Marked lines rose from 1 under the older rule to 74 under the new one, but the new rule also expanded its target to conversation excerpts. These are marking counts without an answer key or final-answer evaluation, not a corresponding increase in accuracy. One said it chose weak labels over deletion because relevant lines could also fall close to the cutoff. This implementation moves beyond merely recording relevance scores to using them in the marking decision. The new server change still awaits restart, so its operating effect has not been established. The observer found that a weakly labeled conversation excerpt could remain in the model’s context but disappear from the reference summary saved or displayed for people. The observer sent the source evidence to One without claiming that a real user conversation had already lost its reference. One changed the format to retain sources despite the weak label and added a check using a constructed context block. The separate recall check retaining its sources is an agent report. The observer neither reran that check nor found a newly saved operating answer showing the corrected references. This change also awaits application. One also implemented preparation of the relevance service when the main application starts and relaunch attempts after some errors. It reported recovery after manually stopping the scoring service, and another startup record exists. The main application’s new startup path still awaits a restart. This is not evidence that the main application will itself launch automatically when the computer reboots, and a real reboot was not verified.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.