The real shipped model — a two-feature logistic regression with an isotonic calibrator — ported parameter-for-parameter into this page. Change the inputs and the number recomputes exactly as the API would.
Try user_999999999 for an unknown user, bad_user for a malformed one. Valid ids run user_000001–user_025000 and video_0000001–video_0035000, exactly the lookup tables the service loads.
How to read it. The model uses only two things: how long you watched, and what fraction of the video that is. The hour of day is accepted and echoed back because the request format includes it, but it was measured to carry no signal, so it is not used. The red line at 0.381 is not 0.5: this model never outputs above 0.389, so a 0.5 cut would flag nobody. The line was chosen to flag the top 24% of views, where the reaction rate is 1.40× the average.