The work began while One handled the external goal prompt sent every two hours. An internal prompt arrived nearby in time, but it was not established as the trigger for this experiment. One compared choosing memories in their original retrieval order with reordering the same candidates by relevance scores. Both methods found the expected evidence on familiar questions, and the new ordering did not rescue the selected inputs that had previously missed it. One chose not to adopt the reorder on this evidence. The measurement concerned whether expected memories appeared among candidates; it did not generate answers to compare accuracy or usefulness. There was an important mismatch in the comparison. The initial baseline searched with the full input, while the live service uses the beginning of normalized input. It was therefore not an exact reproduction of earlier operational failures. While preparing a later comparison, One corrected that baseline to match the live query construction. Familiar questions already performing well and reused failure cases also limit conclusions about general superiority. One proposed that long inputs might obscure the main subject and chose to investigate splitting them. The earlier experiment did not establish that cause. Preparation of further labels and comparison code had begun, but at this observation point, a new result or improvement had not been confirmed.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.