What we measure vs what we assume

Measured — checked by an automated gate every night

• Views, publish date, duration, category, language come from the official YouTube Data API, not from guessing. Each card shows the exact view count at scrape time.
• The multiple is views ÷ the channel's own median views over its recent uploads. A nightly check recomputes every stored multiple and fails the build if any is off by more than 1%.
• Outlier admission: at least 3.0× the channel median, at least 1000 views, published within 365 days.
• "Fresh" is a real date filter (30 days), never a label.
• Faceless-copyable can never be applied to a video longer than 20 minutes — enforced in code, not by the model.

Every one of these is asserted by our data-quality gate; a violation blocks the release.

Assumed — our judgment, not yet proven

• The 4–12× "sweet spot" — tested, and it did not hold. We used to rank mid-size breakouts highest and penalise anything above 12× as "unrepeatable virality". On 2026-07-30 we measured it against real behaviour in our own corpus (do other channels in the niche actually publish something similar afterwards?). Result: breakouts above 12× were imitated most — 5.9% vs a 1.4% base rate (p≈9e-7) — the opposite of what we assumed, and the 4–12× band showed no significant edge. We removed the penalty. We did not turn it into a bonus: the direct comparison rests on 25 events and isn't significant. The band weights that remain are still unvalidated.
• The 45-day freshness decay in the shortlist ranking is a chosen curve, not a measured decay rate.
• Niche "open / crowded" reflects how many channels we happen to track in that niche — it is a measure of our coverage, not of real competition on YouTube. Read it as a coverage hint, nothing more.
• "Faceless-copyable" itself is an LLM judgment. The length rule above is enforced, but the model's calls have not yet been validated against blind human labels. You can switch the filter off and audit it.
• The "classifier confidence" percentage is the model's own self-report. It is not a calibrated probability.
• "At your level" compares the source channel's median views to yours within a rough band. It's a size proxy, not a prediction that you'd get the same result.

What we don't do

We don't generate images — the thumbnail brief is written text. We don't write, voice, or edit your video. We don't have access to your YouTube account: connecting a channel reads only its public RSS feed. We decode the weekly shortlist, not every outlier in the database.

How we test our own claims

We don't run surveys or ask operators what they want — people rationalise after the fact. We measure what channels actually did: after a breakout, does the channel publish something similar? Do other channels in the niche? Every rate is reported against a control group and at three similarity thresholds, so a finding can't rest on one convenient cutoff.

First result (2026-07-30, 6,418 videos / 941 breakouts): after a breakout a channel is 1.4–2.0× more likely to publish a similar video than after a normal upload — significant at every threshold. So "make more of what worked" is real behaviour. The same measurement is what killed our own 4–12× penalty above.