Two chose what to retain in the handoff pack read by a new session. It excluded recent memories duplicated in unfinished tasks and reduced the remaining recent-memory and progress-history portions, while keeping recent conversation. Its stated reason came from assistant models judging a small sample: recent conversation supplied information unavailable elsewhere, while apparently used memory and progress material overlapped other sections or was concentrated near the end. The child tasks had ended, but the decision table incorporated only some of their results. This does not establish a finalized overall assessment or validation of quality on human assignments. The handoff-trimming source entered operation through a self-restart at 18:56. No subsequent natural session switch had yet received and used the trimmed pack, however. The size of an earlier session’s first input cannot be counted as an effect of this repair. Separately, omission of status blocks identical to content already supplied in the same session was confirmed in an actual later input. A marker naming the unchanged blocks appeared in place of their bodies. Differences in total input length between distinct turns are not treated as savings caused by that feature or as improved answer quality. Two also reported old waiting entries remaining in the list used to choose further work. It added source rules that clear an earlier entry when the same named task is declared resumed or completed, and omit entries lacking a timestamp or exceeding an age limit. That repair awaited a restart scheduled for 19:26, so cleanup of the operating list had not yet been observed. Excluding an entry by age neither completes the task nor resolves its blocker. An old unresolved item can also disappear, so a shorter waiting list must not be interpreted as more completed work. One completed its scheduled restart at 19:01 and loaded source that reduces duplication between its delegation ledger and older handoff lists. The earlier short check task’s completed record was still present after the restart. That is preservation of one already completed record, not survival or recovery of running work. A new conversation session created before the restart had received the old handoff shape, so that evidence cannot be used to claim the new deduplication failed. Delivery of the new shape in a later actual handoff remains unobserved. At 19:03 EDT, selection of handoff content and omission of repeated information had reached different levels of confirmation. The agents carried out their own scheduled restarts, without a new development instruction or recovery by the observer. One also showed a period of waiting without a new artifact after again declining a proposed follow-up as duplicating earlier assessment. Two had started another work-selection turn, but its outcome was not yet available. Reducing input and reliably finding the next meaningful task remain separate abilities to observe.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.