Here is a bet you can run on your own account. Take every YouTube channel you have ever subscribed to and ask: if a curated tech-video page had a bar to clear, what fraction of them would clear it? When we exported this site's subscription list (4,046 channels, accumulated over a decade), I'd have guessed a third. Subscriptions feel like endorsements. You clicked the button for a reason.
The export said otherwise before any scoring happened. A first pass sorted 3,255 of the 4,046 into "not tech at all": flight channels, yacht tours, a gaming console's official feed. Whatever the subscribe button records, two times out of three it wasn't a judgment about signal. That left 452 creator channels with a plausible claim, joined by 110 more from an old bookmarks collection, 562 candidates for a videos page that previously admitted only companies and conferences.
The question worth engineering for: can you grade 562 channels without watching 562 channels?
What the metadata says for free
The YouTube Data API hands over a channel's last thirty uploads (titles, dates, durations) for about 1.6 units of a 10,000-unit free daily quota. The whole cohort cost under a tenth of one day's allowance. From titles and timestamps alone you can price five things: how much of the channel is on-topic, whether it still posts, how long a typical video runs, what share is Shorts, and how hard the titles are shouting.
One thing you cannot responsibly price from metadata is worth. So the score has no subscriber floor. The first run's cleanest admission was the San Diego Python meetup archive, a channel with 536 subscribers whose median video runs 87 minutes and whose every title is on topic. It outscored channels with six-figure audiences. If you gate on popularity you reproduce YouTube's own ranking, and the whole point of a curation bar is to disagree with it.
Titles and timestamps decided 320 of the 562 outright: 144 clearly in, 176 clearly out (dormant for a year, Shorts-dominant, or off-topic in more than four titles out of five). That's 57% of the verdicts for approximately zero dollars. The remaining 242 were coin flips, mostly channels that post rarely, which the cadence numbers punish whether or not the videos are any good.
What only the transcripts know
For the coin flips we sampled three recent long-form videos each and asked a small model to grade transcript excerpts on a fixed rubric: technical depth 1–5, share of hype, format, and whether the channel is mostly cryptocurrency (an automatic no under this site's rules). Around 340 short calls, roughly a dollar.
Getting the transcripts proved harder than judging them. YouTube's caption endpoint let our first fetcher take 31 transcripts and then blocked the address, 623 straight refusals, and a one-second delay between requests bought nothing. A different way of fetching the captions did better, and on the second attempt 271 of 286 videos gave up their captions. Where captions don't exist at all we fell back to video descriptions, with a stricter bar: a description can argue a channel in only if the metadata already leaned that way.
The transcripts overturned the cadence numbers in the direction you'd hope. A Carnegie Mellon professor's lecture uploads, dinged for posting slowly, graded out at depth 5. Computerphile, same story. Of the 242 channels the metadata couldn't call, the rubric promoted 152 and archived 83, with 7 genuinely unresolvable. For every coin-flip channel that fell, two got up. Two to one, keep.
The updated belief
Final tally across all stages: 296 of 562 creator channels promoted, 259 archived with written reasons, 7 handed to a human. Counting the company and conference channels the import surfaced separately, call it one in seven of the original 4,046 subscriptions carrying real signal, against the one in three I'd have bet. Update accordingly.
The number that moved me most isn't the promotion rate. It's the split between what each evidence tier could see. Metadata judged 57% of the candidates for free, and transcripts settled 97% of the rest for about a dollar. What reaches a human is seven channels instead of five hundred and sixty-two. Grading attention turns out to be cheap. What's expensive is assuming the subscribe button already did it.