Why timestamps survive chunking better than IDs

A natural first instinct is to give every chunk a numeric ID and cite that. It looks tidy until you reprocess the transcript with a different chunk size, or merge a videoupdate, and every old citation now points at the wrong span. Timestamps tie the citation to the video itself, not to your indexing run.

For an edited re-upload, you treat it as a new video and rebuild the index. That is the same blast radius as switching chunk sizes, but at least your old citations point at a clear "video changed" failure rather than at silently wrong text. In practice, captioned videos almost never get re-edited, so the timestamp strategy holds up across the lifetime of a typical archive.