The input behind the caption mentions in _data/impact/talks.json. Committed, small and few on purpose: keeping them makes that file reproducible, and keeping only windows around the moments SQLancer comes up avoids storing a transcription of somebody else's whole talk.

YouTube serves its caption tracks only to its own player, and for some talks it will not even fill its own transcript panel. So this step is done by a person. Any of these works — the importer takes whichever shape the text arrives in:

1. The transcript panel. Open the talk, expand the description, click Show transcript, then use the panel's ⋮ menu (Toggle timestamps off is not needed) — select the transcript and copy it. Save it to a file.
2. A caption file. yt-dlp --write-auto-subs --sub-lang en --skip-download --sub-format vtt <url> leaves a .vtt beside itself. A .srt works too.
3. The JSON shape this directory holds, if you already have segments.

Then:

python3 -m tools.impact.talks import <file> <talk url>
python3 -m tools.impact.run collect --only talks

The first command prints every mention it found, with its timestamp and how the transcriber heard the name, and writes the windows here. The second folds them into the dataset. Rolling repetition in auto-captions — each cue restating the tail of the one before — is dropped on the way in.

A machine transcription, wrong often enough to matter: it renders "SQLancer" as "SQL answer", "SQL lenser" and "SQLanswer" in the talks collected so far. The records built from it carry the misheard form and a link to the second it was said, and the page says plainly that the words are the transcriber's rather than the speaker's. Where a mention is on a slide and never spoken, no transcript reaches it at all — that is what the frames in assets/images/impact/talks/ are for.