**The most counterintuitive finding is that the usual "fixes" for late-night screen use barely register, while the sheer number of minutes on the phone at bedtime dominates everything.** **What does matter** - Bedtime phone minutes correlate strongly with sleep latency (r ≈ +0.86) and with next-day fatigue (r ≈ +0.71). They correlate negatively with total sleep (r ≈ −0.55). Across 8,500 users, this is by far the strongest behavioural signal (steps 0–5). **What barely matters** - **Blue-light filter:** Filter users (3,976) and non-users (4,524) look almost the same. Average sleep latency is 39.6 vs 41.6 minutes, and average sleep is 6.33 vs 6.21 hours. The correlation of filter use with latency is only −0.057, and with sleep hours +0.039 (step 3, step 5). - **Screen brightness:** Brightness correlates only +0.093 with latency. - **Physical activity:** Activity has essentially zero correlation with sleep hours (+0.004) and with fatigue (−0.029). - **Late caffeine and sleep length:** Post-5pm caffeine has almost no link to total sleep hours (−0.038). It does show up in sleep architecture: it correlates −0.34 with deep sleep and +0.20 with latency. So caffeine appears to change how well people sleep, not how long. **Why it's surprising** A filter, dimmer screen, or exercise would be the obvious things to recommend to reduce the harm of bedtime phone use. In this data they are close to irrelevant next to simple screen minutes. That suggests that cutting phone time, not modifying the screen, is what moves latency and fatigue. **Caveats** - These are correlations on observational data, so none of this proves causation. - Individual phone-minute values are person-identifying, so I used only aggregate correlations. - Sleep hours and snoozes are almost mirror images of each other against fatigue (−0.88 and +0.88). Fatigue seems largely a function of sleep duration, which is itself driven by phone time. The numbers come from the correlation query (steps 0–5) and the gold-table breakdown by `blue_light_filter_active` (step 3).