A real run, with part of it held back
This is the timeline table from one actual aitubenotes run, against a YouTube talk about how AI search decides what to cite. The left column is what was said. The right column is text read off the screen at that moment. Nothing below was written by hand, and the OCR mistakes have been left in.
Loading the sample.
Why is some of it locked?
Because this is the shop window rather than the product. The screen text is sampled across the whole video rather than cut off partway, so what you see is representative of the run instead of just its opening minutes. The heading on each gated column says how much it is holding back.
A run of your own comes back whole: every row, every column, plus the full transcript and the extracted screen text as files you can download.
Why does the screen text matter?
A transcript is made from audio, so anything a presenter showed but did not read out is lost. In this run the speaker says "yada yada yada" while the slide carries the actual sentence. Scroll the table sideways and you can watch it happen: the spoken column goes vague exactly where the screen column gets specific.
Two of those moments are pulled out and explained here, quoted cell for cell from the same run.
What we will not claim
Optical character recognition on video frames is imperfect, as the table shows. Handwriting, dense small text and low-resolution uploads read poorly. Nothing here is reviewed by a person before it reaches you, so the notes carry a line saying a machine wrote them, and anything that matters should be checked against the source. This is one run on one video, and it is not a benchmark.