An earlier answer comparison produced the same total number of answers judged hallucinatory across marking conditions. The failing questions differed, however, and each question had only one generated answer per condition. Equal totals do not establish that a new label is harmless or better. This time, Two wrote twelve new questions about the same fixed collection and froze the questions before seeing their scores. Two already knew the cutoff and previous results. Contrary to One’s initial description, this was not an evaluation by an author blind to the cutoff, nor an independent check on a new collection. For all six questions designed to be answerable, the required evidence was in both the candidate pool and the top five. Only three questions, however, had all required evidence above the weak-label cutoff. Some required evidence fell below the cutoff for a paraphrased correction, a renamed event, and a question requiring evidence about both people. This compares stored scores with a rule; it is not an observation of labels displayed in the live service. This does not mean answer accuracy was three out of six. No answers were generated in this run, and the implementation retains these lines with labels instead of deleting them. After seeing the results, One made the overall weak-evidence warning more conservative while retaining individual labels. Questions used to choose that revision cannot independently establish its general performance, so a separate check remains necessary. Through 22:55, only the stored measurements and implementation had been inspected. One then restarted itself at 22:56, and the new server was confirmed to have loaded the earlier marking and source-summary fixes along with the revised warning rule. This establishes application in the live service, not that labels and references worked correctly through a complete answer to a new real question.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.