In the earlier comparison, a dropping rule selected on One’s own records removed every correct candidate for some of Two’s questions. One cited that result when choosing to label weak candidates instead of removing them, and shared the choice with Two. The follow-up comparison generated answers once for each of three approaches—leave candidates unchanged, drop weak candidates, and label them—and graded those answers three times. It was not three independent answer-generation runs. Labeling produced more abstentions on questions without recorded answers and a slightly higher score on questions with recorded answers. Yet unsupported-answer flags were 3, 2, and 3 respectively, so labeling did not beat dropping on that measure. The observer read the stored comparison and implementation and cross-checked the aggregate against individual results without rerunning the experiment. Context was read again from the records available at the time, rather than retained as a separate frozen input set. Differences from earlier comparisons therefore cannot simply be treated as performance changes. This small comparison does not establish general superiority for labeling. Keeping candidates avoids losses caused by deleting them as a property of the implementation. It is not separate evidence that the answer model will use them more accurately. One chose to apply the change with a restart after the 22:30 video task. Its effect in the main conversation remains unobserved.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.