The corpus

The index is built from a curated list of channels, not from open crawling. A daily run walks those channels, pulls each candidate's automatic captions, aligns them, scores the clip, and writes it to the index.

What gets in

A candidate has to clear every gate:

Gate	Threshold	Why
Captions present	required	Every learning feature is built on word timings. No captions, no clip
Duration	up to 180 s	Short form. Long videos are a different product
Words spoken	at least 25	Filters music, montages and silent clips
Speech density	0.9 words per second and up	A clip where someone actually talks
Safety and topic screening	model reviewed	Curated corpus, usable next to a child account
New vocabulary	at least one word not already saturated	Stops the index filling with the same hundred words

Clips that fail are recorded with a reason rather than forgotten, so a later run does not re-download them and a threshold change can be replayed against them.

What that costs you in expectations

• The corpus is curated, not exhaustive. A word can be common in the language and absent here. GET /words tells you which words are covered before you build a feature that assumes one.
• It grows daily and is not a frozen dataset. Clip counts for a word go up over time.
• It shrinks occasionally, when an uploader deletes a video. Clip ids are stable; the video behind one is not guaranteed forever.

Languages

English is indexed today. Adding a language is a data job, not a code change: a levelled word list, glosses, a channel list, then a corpus run. GET /corpus is the live answer to what is available; nothing in the API assumes English.