Fourteen Reddit searches, one per model name. Stacked, deduped, split, then classified by Haiku and written straight into the ontology. Nothing is staged: this is one run, from the click to the rows.
The clip runs 22 seconds. The first four and the last ten are real time, so you see the run
actually start and actually land; the 170 seconds of grinding in between are compressed hard.
The last thing to appear is both_labelled, 184 rows, which is the point: the run
ends as queryable objects, not as a log line.