Nothing on this page was typed in by hand. Seven jewelry-store YouTube channels are walked with yt-dlp, which also pulls the caption tracks YouTube's own speech recognition has already produced: that is the listening step, and it is what turns hours of dealer talk into text a machine can read. From there a money matcher catches every dollar figure, the words around it decide whether the desk is buying, asking, reporting a sale or refusing, and a watch vocabulary resolves which piece is being priced from references, model names and the nicknames the rooms actually use. Every row is hashed and stamped with the video's publish time, so a re-extract inserts what is new instead of rewriting what was already said.

1. 01

   Harvest

   • yt-dlp
   • channel + video metadata

   7 jewelry-store channels, walked video by video. Publish time comes along, which is what every price is later dated by.

   5,245 videos

2. 02

   Speech to text

   • YouTube ASR captions
   • VTT

   The machine listening step. YouTube's own speech recognition has already transcribed these rooms; we pull those caption tracks rather than re-running audio.

   4,281 transcripts

3. 03

   Money and intent

   • pattern extraction
   • intent lexicons

   Every dollar figure in the text is caught, then the 180 characters around it decide what kind of price it is: a buy, an ask, a sold, or a hard pass.

   6,106 spoken prices

4. 04

   Which watch

   • reference vocabulary
   • nickname map

   126710BLRO, Speedmaster, Pepsi, Batman, Hulk. References, model names and the nicknames dealers actually use all resolve to the same piece.

   30 named pieces

5. 05

   Date-stamp and dedupe

   • fingerprint hash
   • append-only

   Each row is hashed and stamped with the video's publish time. Re-extracting inserts what is new and leaves history alone, so the book never quietly rewrites itself.

   2 passes kept

A number in the room is not a price. Retail, a guide, an auction result, and “you could list it for two million” stay on the tape and off the book. So does a caption that hears “fifteen five” as $155,000. The buy and the ask are only lines where a desk is actually bidding or asking, and a print that sits far outside that model’s own range does not set the headline.

What is not running yet: a language-model pass over the windows the pattern layer flags as ambiguous, to read intent where a lexicon cannot. Every figure on this page came out of the deterministic pass above, which is why the board admits its own noise rather than smoothing it.