**The most counterintuitive finding is that the usual sleep-hygiene fixes barely register in this data. Blue-light filters, screen brightness, late caffeine and exercise show almost no link to sleep duration, while the sheer minutes of bedtime phone use dominate.** These are all linear correlations across 8,500 users from the correlation queries I ran (steps 0 and 1). The data is observational, so none of this shows cause and effect. **What the data says about phone time:** - Bedtime phone minutes correlate very strongly with sleep latency (r = 0.86) and next-day fatigue (r = 0.71). - They correlate negatively with total sleep hours (r = -0.55). **What the common remedies show:** - **Blue-light filter:** The filter is essentially uncorrelated with sleep latency (r = -0.057) and total sleep hours (r = +0.039). It is the most widely recommended fix, yet users with it active look almost the same as users without it. - **Screen brightness:** It has a very weak link to sleep latency (r = 0.09). Dimming the screen looks far less important than simply putting the phone down. - **Caffeine after 5pm:** It has no linear relationship with total sleep hours (r = -0.038) and only a modest one with latency (r = 0.20). Its clearest footprint is on deep sleep (r = -0.34), so it seems to hurt sleep quality rather than quantity. - **Physical activity:** It is essentially unrelated to sleep hours (r = 0.004) and fatigue (r = -0.029), despite being a staple of health advice. **Why it matters:** The pattern points to one dominant driver, which is how long people spend on the phone in bed. If so, a product that adds a blue-light filter or a caffeine cutoff may do little unless it also reduces phone time. Fatigue also tracks sleep hours almost one-to-one (r = -0.88) and alarm snoozes (r = 0.88), so it looks like a downstream symptom rather than an independent outcome. **Caveat:** Correlations only capture linear effects. A filter or caffeine effect that applies only to some subgroups, such as certain chronotypes or heavy phone users, could be hidden. The pre-aggregated chronotype and filter breakdowns would be the next place to check.