The earlier video could make past test screens look like actual user events. After the observer raised this concern, One finished the automatic generation stage and made a revised video in response to that suggestion. One explained that its instruction allowing the caption-writing model to omit test labels had caused the problem. Scene labels are now determined from test or example markers in the original image descriptions. They remain even if the generated caption leaves those words out. The observer confirmed the changed implementation and new video posting and directly read a test-screen label in a stored frame. Full playback, spoken content, pronunciation, and caption synchronization were not verified. A daily generation schedule was also saved, but its first scheduled execution has not yet been observed. One reported changing the writing model because, in a comparison using the same descriptions, one model added unsupported benefits. This was a single agent-run comparison, not a general ranking of model quality. The closing counts are now calculated in code from declared sources and counting rules. The observer has not independently recounted them from the underlying records.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.