Examples today
40
179 labeled by catalog ratio, 40 with full features
Examples needed
500+
roughly 20 reform channels for two months of uploads, or any partner archive
Training cost
~$0
minutes of laptop compute; the data is the only expensive ingredient
1
CollectPoller running hourlyThe velocity poller watches every new upload's first 72 hours across the reform channels. Each curve is one labeled experiment.
2
LabelRule writtenThe verdict comes from the curve, not the lifetime count: an exploit inflection inside 72 hours is a pass, a plateau is a kill.
3
ExtractRunning todayThe existing pipeline already produces the features per video for about $0.04: transcript with word timing, pace, pauses, loudness, pitch, captions, hook, call to action. Delivery labels (tone, facial read) are one prompt away in the same pass.
4
TrainAn afternoon once the data existsGradient-boosted trees or plain logistic regression on the binary verdict; class weights for the rare positives; probability calibration so 0.7 means 70%.
5
ValidateDesignedLeave one channel out entirely and test on it, so the model must generalize across audiences rather than memorize FairVote. Report AUC and lift in the top decile, nothing fancier.
6
Ship as a priorUI exists on /break-outThe model's probability becomes the pre-publish iWAR estimate shown with its inputs, the way every PAX2 number shows its math. It advises which take to post first; the live curve then renders the real verdict.
1.Any weight on a single feature: with a handful of pushed videos, one lucky caption can look like a law. The lift table stays a table of hints until step 1 runs.
2.Anything about persuasion. This whole page predicts travel; whether a traveling video changes a mind is a different experiment with a different control group.
3.Causation. Catalogs are observational; the Upworthy archive is the only randomized evidence in the building, and it only covers wording.